Use artificial delays only with emulated hardware.

Do reduction more frequently to avoid a rare case of failure during
reduction of a too large quadratic form.

Update init_hw to use standard Open().

Add stubs for I2C read/write and setting PLL frequency.

Take into account that elapsed time may be zero.

The hardware VDF client connects to the timelord and runs VDFs using
challenges (initial values) provided by the timelord. All 3 VDFs are run
in the same process (unlike the original VDF client). Proof generation
is not supported yet and will be added in a future version.

Prover class could now be used from other compilation units.

Clear work queue when finishing a VDF run and not waiting for a proof.

Generating durations with a seed allows better reproducibility.

This allows Valgrind to schedule other threads to run, which doesn't
happen without the sleep.

Use values that produced errors when testing VDF client.

It may happen that a new job is started on the same VDF engine before
the existing one was able to finish. We must wait for it to complete.

This resembles how real timelord communicates with VDF client.

Move timestamp-related functions to hw_util.hpp.

Read and log ASIC temperature every 2000 reads of status registers
(every 200 for emulator).

Notify timelord and disable VDF engines before shutting down.

Emulate normal timelord behavior.

HW VDF client can run up to 2 background threads computing proofs per
VDF engine.

An error has been observed where the VDF client connects to timelord
emulator immediately after closing the previous connection, but before
the timelord emulator could free the VDF slot. As a result, the
subsequent connection was being refused due to a lack of available
VDF slots.

The standard abs() function defined in stdlib.h takes and returns an
'int' argument (32-bit), but we expect a 64-bit integer. When another
header file, gpu_integer.h, was also included, it provided a different
function named abs(), which was returning a 64-bit integer. Builds of
some programs such as hw_vdf_client do not include gpu_integer.h, and
this caused incorrect form reduction in very rare cases, so the problem
went unnoticed for a long time.

Change to the standard labs() function to get the correct behavior
regardless of extra include files.

Compute proof checkpoints every million iterations, use them to return
n-Wesolowski proofs with multiple segments in hardware VDF client.

Using checkpoints reduces delay after a proof is requested since we only
need to compute the proof from the previous checkpoint.

When running VDF hardware emulator, a single bad VDF value may appear
when switching to a new discriminant.

Only warn about work queue size for VDF auxiliary threads when a
threshold is reached; increase or decrease thresholds when a higher or
lower threshold is reached, respectively.

Reduce proof delay by having all proof segments complete when the last
segment is complete. Do not create proof checkpoints at regular
intervals, instead use the 2-weso proof technique similar to one in
software VDF client.

selected through FT4222H OTP.

… hw_vdf

Set number of proof threads to 7/8 of the total number of aux threads.

When a proof is requested for lower iteration count than what is already
queued, remove existing queued proofs and insert the new proof request.
Previously such proof requests were waiting for existing proof requests'
completion unnecessarily.