perf(lexer): gate keywordStart on the next byte for i/c tokens

[BridgeAR]

Summary

The main character loop calls keywordStart() (and through it a second boundary helper) on every i and c token it reaches. On the angular/d3/rollup/magic-string corpus the byte after the token rules it out ~98% of the time for i and ~95% for c, so those calls are almost always wasted. Checking *(pos + 1) against the keyword's second letter (m for import, l for class) first short-circuits the calls with a single load+compare. memcmp already requires that same byte, so the guard is behaviour-preserving.

Why

The per-character loop is the dominant cost (it runs once per source code unit; the boundary helpers are called <0.05x per code unit), so cutting a function call out of the two hottest keyword cases moves the needle where micro-tuning the helpers themselves does not.

Measured on the same corpus the README benchmarks (~3M code units, Node 24 / V8, 150 iters x 13 trials, drop best+worst, run against the built wasm):

• parse time: -1.4% mean (every trial negative), parse output byte-identical
• wasm: -229 bytes / -129 instructions

Drive-by

• isBrOrWs: write the 9..13 range as (char16_t)(c - 9) < 5, one unsigned compare instead of two. The wasm backend keeps the shorter form at every inlined copy of the helper; this accounts for most of the byte/instruction reduction above.

[guybedford]

Looks good thanks, although seems wasteful to do the double compare if we can avoid it?