perf(js): stop allocating buffers for each F32 bit cast

[oxura]

Fixes #1638.

The JS bit casts share a four-byte buffer instead of allocating one for every call. They reuse the cast functions from bend.ts, with no second implementation or endian flags.

Kept the round-trip table for signed zero, subnormals, finite bounds and infinities; removed the unrelated lib.bits(0.75) call. No new NaN-payload guarantee.

Local generated-module smoke: 0 / 16777216. Repo gate: 54/54. The maintainer ran the full gates on the previous head; I haven’t rerun the private gates locally.

[Lorenzobattistela]

Thanks! The speedup is real and all gates are green. I pushed a commit that drops the GUIDE edit; we don't want that change in the guide. Three small things before we merge:

1. Don't duplicate the bend.ts helpers. bend.ts already has these exact casts: F32_VIEW, f32_to_bits and f32_from_bits. comp.ts already splices ${Bend.f32_round} into the JS runtime, so reuse the bend.ts versions the same way instead of keeping a second copy in comp.ts.
2. Drop the , true arguments. Each cast writes and reads with the same byte order, so the little-endian flag changes nothing. Big-endian, the default that bend.ts uses, gives the same bits. The four arguments just cost ttok in comp.ts.
3. Remove lib.bits(0.75) from the test loop. It can't catch a scratch-reuse bug: value is already a JS number before that call, and lib.bits(value) rewrites the buffer before reading it. The round-trip table on its own is the useful part.

[oxura]

This is JS-only performance work, so I’m withdrawing it following the reference-runtime policy in #1644.

[oxura]

I closed this too quickly after reading #1644 — sorry. Reopening it and taking care of these three points.

[oxura]

Done — comp.ts now splices in the two bend.ts helpers and uses their default byte order. Dropped the extra test call. The exported-module smoke passes (0 / 16777216), and the repo gate is 54/54.