When backtraces are disabled (due to their large performance overhead),
we can still capture the `Location::caller()`, which is quite cheap.
This is very useful for rare panics in CI, which can be hard to reproduce.

… add `#[track_caller]`s when `debug_assertions`.

…ounds error messages (#980)

When backtraces are disabled (due to their large performance overhead),
we can still capture the `Location::caller()`, which is quite cheap.
This is very useful for rare panics in CI, which can be hard to
reproduce.

This also makes the out-of-bounds error messages more informative (they
say which thing exactly is out of bounds and what the bound was, like
`std`).

…elds.

…elds (#978)

I found these by running the argon tests in debug mode (not just
`debug_assertions`), as these `libstd` assertions are only run in actual
debug mode.

…s` a method.

…_dav1d`, as this is the only callsite and we'll need to store stuff on the stack.

… there are safer wrapper methods.

…ough the callstack.

… side of the `fn` ptr boundary).

…ntMutArcSlice<refmvs_temporal_block>>`s.

…vs_pool` now that `mvs`s have been `Arc`ified.

…l` now that `cdf`s have been `Arc`ified.

This fixes a bug from #971 where I was overslicing. It wasn't caught by
the normal tests, only the argon tests that only ran on `main` once I
merged.

`fn splat_mv` is already a method, and `fn rav1d_refmvs_save_tmvs` was
already like a method, but as a free `fn`. So this just makes them all
methods on `Rav1dRefmvsDspContext`, which helps things for `fn
load_tmvs` as I make it and the `mvs` fields safe.

…ough the callstack (#983)

…tMutArcSlice<refmvs_temporal_block>>`s (#984)

* Fixes `{,ref_}mvs{,_ref}` fields of #713.
* Fixes `refmvs{,_pool}` fields of #641.

This was already `Arc`ified in a previous PR, so this removes the now
unused pool.

…ust mostly nothing if asm is disabled.

…efault.

…ng `&'static` refs in `Rav1dContext`.

…ce code size.

…e `&'static`s (#986)

* Fixes `dsp` field of #713.
* Fixes all `fn *dsp_init`s of #862.
* Improves DSP initialization code size of #809.

This
* Makes all `fn *dsp_init*`s safe.
* Changes all `fn *dsp_init*`s into `fn *DSPContext::new`s that directly
initialize the type without `unsafe`ly zero initializing it first. This
is done by directly initializing in `fn default`, which is for the
fallback `fn`s, and then updating the `fn` ptrs in the `fn
init_{x86,arm}`s. These `fn`s are also all `const` now, as that was
trivial.
* Adds `wrap_fn_ptr!` and `enum_map!`s to the array fields of
`Rav1dMCDSPContext`, since I needed a default value to initialize the
arrays directly if I want it the indices to be named (rather than a
normal array, whose order is easy to mix up). `wrap_fn_ptr!` provides
that default value, which is then optimized out.
* Replaces the `Rav1dContext::dsp` array, which is manually lazily
initialized, with `fn Rav1dDSPContext::get`, which uses `OnceLock` for
lazy initialization.
* Makes the `Rav1dFrameData::dsp` raw ptr into a `&'static` now that we
lazily initialize it with `OnceLock`.

This also dramatically reduces the code size of DSP initialization (in
KiB):

| Target                          | Before | After |
| ------------------------------- | ------ | ----- |
|      `x86_64-unknown-linux-gnu` | 192.5  | 95.6  |
|        `i686-unknown-linux-gnu` | 106.5  | 39.6  |
|     `aarch64-unknown-linux-gnu` |  88.2  | 15.3  |
| `armv7-unknown-linux-gnueabihf` | 103.8  | 17.0  |

This is measured before using `cargo bloat --filter 'dsp_init'` and
after using `cargo bloat --filter 'DSPContext::new'`.