perf(decode): inline VLQ segment parser and skip capacity pre-pass

[Boshen]

Summary

Two micro-optimizations to `decode_mapping`:

1. Skip the upfront `,` / `;` counting pass. It was used to pre-size the token `Vec` to an exact upper bound; replace with a `mapping.len() / 3` heuristic. Realistic sourcemap segments are 3-5 bytes, so this slightly over-allocates without paying for an extra full scan of the input.

2. Inline `parse_vlq_segment_into` into the outer match arm and use a local `local_cursor` that's copied to/from the surrounding cursor at segment boundaries. With cursor state and the `nums` array visible at the function scope the optimizer can keep them in registers across iterations, and the function-call + reborrow overhead per segment is gone.

`get_unchecked` is used for the `mapping[cursor]` reads — every load is preceded by an explicit length check, so the bounds checks are already provably redundant.

Behavior, error variants, and the public API are unchanged.

Numbers (standalone vs main, no other PRs applied)

Small improvement, mostly under noise floor for the on-disk 3 KB fixtures. The win shows up best on the synthesized large fixture in #328:

benchmark	before	after	Δ
parse/real_large	3.41 µs	3.34 µs	~−2%

Most parse savings come from #329 (borrowed deserialization); this PR is a small additional improvement that's primarily about removing redundant work.

[codspeed-hq]

Merging this PR will not alter performance

⚡ 3 improved benchmarks
❌ 4 regressed benchmarks
✅ 9 untouched benchmarks
⏩ 5 skipped benchmarks¹

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
❌	build_single	7 µs	7.1 µs	-1.64%
❌	lookup_table[real_small]	1.3 µs	1.4 µs	-2.13%
⚡	parse[real_large]	50.5 µs	48 µs	+5.18%
⚡	parse[real_xlarge]	1.4 ms	1.3 ms	+6.3%
⚡	serialize[real_medium]	5.1 µs	5 µs	+1.16%
❌	from_json_string_inline	14.1 µs	14.4 µs	-1.66%
❌	parse[real_small]	11.6 µs	12.1 µs	-3.8%

Tip

Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing perf/decode-inline-vlq (eff0ff4) with main (db883f9)}

Footnotes

1. 5 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩