tests: Add criterion for basic benchmarking

[Herschel]

Add criterion with some basic benchmark SWFs, and refactor some of the test code to support this.

criterion seems geared towards smaller benchmarks, and these benchmarks are already a little chunky -- it's possible something like bencher might be more suitable.

TODO, probably future PRs:

• Benchmark AVM more directly
  • It'd be nice to benchmark AVM code directly instead of running the entire frame loop & player machinery.
  • Trickier with AVM1 because it's so coupled to the timeline and displaylist.
  • This should be good enough for comparisons for now.
• TestBuilder could become a more general PlayerBuilder available in core.
• Get rid of the giant list of tests in regression_tests.rs and instead generate tests via directory listing similar to the benchmarks
  • Tests that need a specific config (approx, frame count, etc.) can be specified, otherwise can be omitted.
  • I think this needs to be done via procmacro or build.rs for tests.
• Add some AVM2 and general playback benchmarks.

[adrian17]

Tests that need a specific config (approx, frame count, etc.)

Could also put test-specific configs as special comments in output.txt. (I'm thinking of something like LLVM's tests.)

But also... are there cases where we want different frame count than the number of frames in the test SWF itself?

[adrian17]

Also, let's maybe rename the benches to make it explicit they test AVM1?

[Herschel]

are there cases where we want different frame count than the number of frames in the test SWF itself?

Quite a few, if you look through the tests (gotoAndStops, looping anims, timers, ...). But using the # of frames of the main timeline is a sensible default that I'd like to add.

Could also put test-specific configs as special comments in output.txt. (I'm thinking of something like LLVM's tests.)

I started working towards a more general test + SWF compiler over here, and I figure these settings can be a part of that syntax. But in the meantime doing some YAML front matter in output.txt would work nicely.

[Herschel]

Currently the results seem to vary quite a lot across runs, so I'm going to try to narrow it down to run the AVM code more directly.

[adrian17]

large_array_init feels small enough that it might be beneficial to even wrap the array creation in an ActionScript loop, to minimize relative startup overhead compared to actual AS runtime.

[zamazan4ik]

Hey, @Herschel! What is the status of this PR?