Create a reference decode speed benchmark page

[jamsinclair]

Inspiration came from recent references to decode speeds in #44 and #43.

We could likely create a simple benchmark script and HTML page that renders the results. This library is not focused on speed, but it would be convenient to compare speeds between decoding with browser image and canvas APIs vs the WASM decode speed.

[CodeF53]

When making this I would use OffscreenCanvas for the VanillaJS side of the benchmark, as I have found using it tends to have slightly faster results.

[ckellet]

There's a harness for this in a fork, if it's a useful starting point: bench/.

It reports wall time, output size and SSIM together, on the grounds that a change which shrinks output by degrading the image isn't an improvement and byte counts alone can't tell the difference. Runs against the built packages and writes JSON so two branches can be diffed.

Two things it does that I'd suggest any version of this keeps, because both cost me a false conclusion first:

• Warm up before timing. WebAssembly starts in V8's baseline compiler and tiers up to TurboFan after a few executions. Timing from the first call mixes two compilers into one distribution — on its own that manufactured an apparent 26% regression for me.
• Refuse to record on a loaded machine. It records load average with the results and won't write a baseline file when the machine is oversubscribed. Sizes and SSIM stay valid under load; wall time doesn't.

There's also an ab.mjs that interleaves two builds trial-by-trial in one process, for when a quiet machine isn't available — its noise floor measures at 0.1% comparing a build against itself.

Source image is generated from a seeded PRNG rather than committed, so it's reproducible without carrying megabytes in git.