bench: engine load tools, a worklet over-budget counter, and an instruction gate

[mishan]

The worklet times every process() call against the quantum it rendered
(Date.now, the only clock in its scope: 1 ms steps, so `over budget' is
a reading of 3 ms or more at 48 kHz) and posts the count over budget,
the slowest call and a histogram by whole milliseconds with each tape,
plus whether the clock is too coarse to trust. host.js's synth.quanta()
keeps the totals and the last ~2.7 s apart; the jam numbers panel shows
both. A tape from an older worklet without the fields is delivered
uncounted.

bench.mjs plays each piece the way the page does -- samples, default
patches for empty channels, seed 1 for pieces with none -- so every
shipped piece is timed, and says which channels were timed as silence.
--json writes one document: per piece and chord p50/p99/p99.9/max per
quantum against the budget, failures as entries, and the build timed
(its path and wasm hash) beside the script's commit. --only and
--seconds are checked; a run that times nothing fails.

dspbench gains -j (JSON), -l and -r (window length and rate), the
chord's every window timed (nearest-rank p99 over how many windows),
and -s/-h in milliseconds, so play loads compare across window lengths.

worklettime.mjs plays pieces on the solo page in a real AudioContext
and reports over-budget calls per minute, the slowest call, p99 and the
audio clock against the wall clock; a page error fails the piece.

instrcount: dspbench on four instruments under cachegrind, against a
baseline keyed by compiler, flags, LTO, glibc and Valgrind, failing
over x1.10 or under x0.98 so a speedup is recorded with --update in the
commit that made it. CI's toolchain is committed; a desktop's counts
live in its build. CI pins the linux job to ubuntu-24.04 so an image
change is a commit that records new counts, and prints the margin.