perf: stop the code tab forking git per tree entry

[espadonne]

Phase 3 (CPU) of the availability campaign: the repo home / code tab is
the most-crawled page on the site and forked git O(entries) per
anonymous request.

• de3ab45a — every git subprocess in internal/repos/git now goes
  through one helper that counts forks (repogit.ForkCount()) and can
  carry process-wide policy. Read-only calls on the code-tab paths get
  a 30 s deadline (repogit.ReadTimeout); blob streaming and
  pack/transport are excluded.
• 909ba217 — repogit.EntryLastCommits: one streamed
  git log --name-only walk per directory, killed mid-stream once
  every entry resolves, bounded at 2,000 commits. Entries the walk
  can't resolve fall back to the old per-path git log -1, so
  rendered output is unchanged.
• 6c0a803b — internal/web/handlers/repo/treecache: OID-keyed LRU +
  singleflight over the last-commit map, rev-list --count, the
  ls-tree -r language aggregate and the log -n 500 contributor
  tally. 2,048 / 512 entries, 10 min TTL. Invalidation is structural —
  a push changes the commit OID, so it changes the key; no hook to
  forget.
• eeba23c1 — wires it through repo.Deps.TreeCache, built once in
  internal/web/repo_wiring.go (same shape as httpcache.PageCache).
• 98219e85 — handler-level fork-count regression test.
• a6eb19f8 — docs/internal/caching.md, docs/internal/code-tab.md,
  sitrep item ticked.

Fork counts (measured, cold anonymous root-tree render)

Directory	Before	After (cold)	After (warm cache)
6 entries	15	10	6
81 entries	90	10	6

O(N) → O(1) in entry count. "Before" measured on this branch with the
walk disabled, so it is the same fixture and the same counter.

Tests

• internal/repos/git — single-walk resolver: root + nested dirs,
  renames, never-touched entries, walk-bound stragglers, parity
  against per-path git log -1, fork count constant at 5 and 60
  entries, bad ref
• internal/web/handlers/repo/treecache — hit/miss, OID-change
  invalidation, key independence, TTL, LRU eviction, errors not
  cached, singleflight collapse, nil receiver
• internal/web/handlers/repo — TestCodeTree_*: forks constant
  in entry count + ceiling, warm cache drops forks with identical
  body, post-push render is not stale
• make ci green end to end against Postgres 16 (golangci-lint
  2.13 + gofumpt, all boundary/spdx/unused/migration/systemd lints,
  verify-api-docs, full go test ./..., build), rebased onto
  current trunk