Moves the library off of Faraday and over onto the standard library's
built-in `Net::HTTP` module. The upside of the transition is that we
break away from a few dependencies that have caused us a fair bit of
trouble in the past, the downside is that we need more of our own code
to do things (although surprisingly, not that much more).

The biggest new pieces are:

* `ConnectionManager`: A per-thread class that manages a connection to
  each Stripe infrastructure URL (like `api.stripe.com`,
  `connect.stripe.com`, etc.) so that we can reuse them between
  requests. It's also responsible for setting up and configuring new
  `Net::HTTP` connections, which is a little more heavyweight
  code-wise compared to other libraries. All of this could have lived in
  `StripeClient`, but I extracted it because that class has gotten so
  big.

* `MultipartEncoder`: A class that does multipart form encoding for file
  uploads. Unfortunately, Ruby doesn't bundle anything like this. I
  built this by referencing the Go implementation because the original
  RFC is not very detailed or well-written. I also made sure that it was
  behaving similarly to our other custom implementations like
  stripe-node, and that it can really upload a file outside the test
  suite.

There's some risk here in that it's easy to miss something across one of
these big transitions. I've tried to test out various error cases
through tests, but also by leaving scripts running as I terminate my
network connection and bring it back. That said, we'd certainly release
on a major version bump because some of the interface (like setting
`Stripe.default_client`) changes.

Drops support for Ruby 2.1 (EOL March 31, 2017) and 2.2 (EOL March 31,
2018). They're removed from `.travis.yml` and the gemspec and RuboCop
configuration have also been updated to the new lower bound.

Most of the diff here are minor updates to styling as required by
RuboCop:

* String literals are frozen by default, so the `.freeze` we had
  everywhere is now considered redundant.

* We can now use Ruby 1.9 style hash syntax with string keys like `{
  "foo": "bar" }`.

* Converted a few heredocs over to use squiggly (leading whitespace
  removed) syntax.

As discussed in Slack, I didn't drop support for Ruby 2.3 (EOL March 31,
2019) as we still have quite a few users on it. As far as I know
dropping it doesn't get us access to any major syntax improvements or
anything, so it's probably not a big deal.

Makes the `code` parameter on `CardError` named instead of positional.
This makes it more consistent with the rest of the constructor's
parameters and makes instantiating `CardError` from `StripeClient`
cleaner.

This is a minor breaking change so we're aiming to release it for the
next major version of stripe-ruby.

Tweaks the retry logic to be a little more like stripe-node's. In
particular, we also retry under these conditions:

* If we receive a 500 on a non-`POST` request.
* If we receive a 503.

I made it slightly different from stripe-node which checks for a 500
with `>= 500`. I don't really like that -- if we want to retry specific
status codes we should be explicit about it.

We're actively re-examining ways on how to make it easier for clients to
figure out when to retry right now, but I figure V5 is a good time to
tweak this because the modifications change the method signature of
`should_retry?` slightly, and it's technically a public method.

I messed up in #828 by (1) accidentally flipping the comparison against
`:post` when checking whether to retry on 500, and (2) forgetting to
write new tests for the condition, which is how (1) got through.

This patch fixes both those problems.

I noticed that we had a couple of other deprecated methods on `Stripe`
and `StripeObject` that have been around for a long time. May as well
get rid of them too -- luckily they were using `Gem::Deprecate` so
they've been producing annoying deprecated warnings for quite a while
now.

Adds a few basic features around connection and connection manager
management:

* `clear` on connection manager, which calls `finish` on each active
  connection and then disposes of it.

* A centralized cross-thread tracking system for connection managers in
  `StripeClient` and `clear_all_connection_managers` which clears all
  known connection managers across all threads in a thread-safe way.

The addition of these allow us to modify the implementation of some of
our configuration on `Stripe` so that it can reset all currently open
connections when its value changes.

This fixes a currently problem with the library whereby certain
configuration must be set before the first request or it remains fixed
on any open connections. For example, if `Stripe.proxy` is set after a
request is made from the library, it has no effect because the proxy
must have been set when the connection was originally being initialized.

The impetus for getting this out is that I noticed that we will need
this internally in a few places when we're upgrading to stripe-ruby V5.
Those spots used to be able to hack around the unavailability of this
feature by just accessing the Faraday connection directly and resetting
state on it, but in V5 `StripeClient#conn` is gone, and that's no longer
possible.

I ended up having to relax the maximum method line length in a few
previous PRs, so I wanted to try one more cleanup pass in
`execute_request` to see if I could get it back at all.

The answer was "not by much" (without reducing clarity), but I found a
few places that could be tweaked. Unfortunately, ~50 lines is probably
the "right" length for this method in that you _could_ extract it
further, but you'd end up passing huge amounts of state all over the
place in method parameters, and it really wouldn't look that good.

This is largely just another cleanup patch, but does a couple main
things:

* Hoists the `last_response` value into thread state. This is a very
  minor nicety, but effectively makes `StripeClient` fully thread-safe,
  which seems like a minor nicety. Two calls to `#request` to the same
  `StripeObject` can now be executed on two different threads and their
  results won't interfere with each other.

* Moves state off one-off `Thread.current` keys and into a single one
  for the whole client which stores a new simple type of record called
  `ThreadContext`. Again, this doesn't change much, but adds some minor
  type safety and lets us document each field we expect to have in a
  thread's context.