perf: parallel blame, parallel link checks, overlapped startup

[laulauland]

Three additional concurrency layers in src/commands/lint.zig, all stacked on the per-doc Io.Group parallelism from #24. Fully documented in docs/CONCURRENCY.md.

Changes

Layer 2: per-binding parallel blame (io.async). checkBinding no longer shells out to git log inline. A new phase in checkOneDocInner spawns one io.async(vcs.getLatestBlameInfo, …) future per changed_after_baseline stale anchor, then awaits all of them and fills in the blame info on the anchor rows. A doc with 5 stale anchors previously ran 5 sequential git log calls; now they run concurrently on the thread pool.

Note: blame is fired after staleness is confirmed, not speculatively. Speculative blame on fresh anchors would cost ~20 ms per fresh anchor (typical git log -1 latency) to save ~15 ms per stale one — net negative in the common case where most anchors are fresh.

Layer 3: per-link parallel existence checks (Io.Group). checkDocLinks now spawns one task per markdown link, each doing realPathFileAlloc + accessAbsolute. Results fill pre-allocated ?JsonLinkRow slots and are appended in doc-order after group.await. Io.Batch was evaluated and rejected because its Operation union doesn't include stat/access ops in Zig 0.16.

Layer 4: overlapped startup (io.async). vcs.getRepoIdentity now runs as an io.async future in parallel with discoverDocGroups. Two independent git subprocesses execute concurrently. Small absolute win, but free.

Testing

• zig build passes
• zig build test passes: 63 tests, 0 failures
• drift check output byte-identical to sequential baseline (same docs, same blame info, same sort order)
• Local timing on drift's repo: ~0.14 s (layer 1 only) → ~0.11 s (layers 1–4). The layer-1 win from feat: parallelize per-doc check loop with Io.Group #24 dominates; layers 2–4 matter more in specific shapes (many stale anchors per doc, many cross-links per doc, tighter startup budgets).

Docs

docs/CONCURRENCY.md replaces the "proposed future work" sections with the actual implementation, keeps the rationale for why speculative blame was rejected, and notes one remaining proposal (persistent GitCatFile request/response pipelining) that doesn't match any current use case.