Perf: avoid using `chunks_exact*` when size is known statically

[CrazyboyQCD]

chunks_exact* methods give runtime slice while as_chunks* give compile time length array to reduce possible bounds checks.

[robobun]

✅ Done → #31415

Converted every static-size chunks_exact* call site in the repo to as_chunks* so the chunk length is a compile-time array (&[[T; N]]) and the per-element bounds checks get elided:

• ast/char_freq.rs — 32-byte histogram chunks
• runtime/image/codecs.rs — 4-byte RGBA pixels
• runtime/webcore/encoding.rs — 2-byte Latin-1→UTF-16 widening
• bun_core/lib.rs (copy_ascii_prefix) — the 8-byte SWAR ASCII fast path (also drops a fallible try_into)
• http_jsc/websocket_client/WebSocketUpgradeClient.rs — 2-element header pairs
• install/lockfile/Package.rs — the Meta/Bin validation loops (concrete types → const stride + const field offsets)
• scripts/verify-baseline-static — the aarch64 scanner’s 4-byte instruction-word loop

The only remaining chunks_exact is the Package.rs Resolution loop: its stride is size_of::<ResolutionType<SemverIntType>>(), which depends on a generic type parameter and would need the unstable generic_const_exprs.

Semantics-preserving; covered by new tests for the Latin-1→UTF-16 widening (buffer-utf16), the lockfile field loader (bun-lockb-field-validation), and the aarch64 scanner, plus existing suites including the TextEncoder ASCII-fast-path boundary tests.