chore(bench): add parse-phase benchmark

[luojiyin1987]

Summary

• Add scripts/bench-parse.mjs.
• Add the bench:parse script.
• Add a Node 22 CI smoke step.
• No parser runtime change.

Phases

The benchmark measures four phases on the same input:

phase	call
A	parseMd
B	bare fromMarkdown + parser extensions
B2	B + recordingExtension
C	parseMdWithSourceMap

Derived values: A-B (remark wrapper), B2-B (recording), C-B2 (post-parse source-map setup / indexing), C-B (total source-map increment).
indexNode is likely most of C-B2, but this harness does not time it alone.
C-B2 also covers RecordingState / WeakMap creation, the recordingExtension(state) call, parse-time snapshots, and the sourceMap object with its closures.

Each measured sample runs in its own child process.
The parent only aggregates.
A single process builds a large multiline AST with heavy GC churn, so an in-process median flips between runs.

Findings (256 KiB, local)

• multiline-hmd: base parse is almost the whole cost.
  recording and post-parse setup stay near zero.
• mixed-markdown: base parse is about 85 percent.
  recording is about 190 ms.
  post-parse setup is about 26 ms.

The benchmark does not identify one slow micromark tokenizer yet.

Smoke design

• Check that every phase runs and returns a finite wall time.
• Check AST parity across A, B, B2, and C.
• Check 4x input growth with a ratio budget.
• Do not check absolute wall time.

Note: multiline-hmd grows super-linearly at large sizes.
The min ratio from 64 KiB to 256 KiB is about 15x on a 4x input.
The smoke uses 16 KiB to 64 KiB and a 10x budget instead.
This keeps the check stable and still catches a gross regression.

Validation

• pnpm run lint
• pnpm run build
• pnpm run test
• node scripts/bench-parse.mjs --smoke
• node scripts/bench-parse.mjs --shapes single-line,entity-dense,escape-dense,many-paragraphs,large-code-block --sizes 65536 --runs 1

Closes #125