CS-11029: in-process inflight dedup for #moduleCache transpile

[lukemelia]

Stacks on #4750 (CS-11028). Review/merge that first.

Summary

• Adds Realm.#inFlightTranspiles — a Map<LocalPath, Promise<…>> keyed by local path. The first caller to miss #moduleCache installs a pending entry; concurrent same-path callers await the same promise instead of each running babel from scratch (50–500 ms per duplicate transpile saved).
• Refactors loadModuleFromDisk to delegate the materialize + transpile step to #transpileModuleDeduped / #materializeAndTranspile. Public behavior unchanged — same response shape, same etag, same error handling.
• Identity-checked .finally cleanup mirrors CachingDefinitionLookup.#inFlight (CS-10947) — a newer pending entry installed after an invalidate drop is preserved when an older promise eventually settles.
• The #dropModuleCacheEntry / #dropAllModuleCacheEntries helpers from CS-11028 now also drop the in-flight entry, so every invalidation site (writeMany, delete, file-watcher callback, executable-invalidation cascade, full-index clear, etc.) clears in-flight too. Existing waiters on a dropped promise still receive its bytes — their requests preceded the invalidate, so pre-invalidation bytes are correct for them. New callers don't join the stale transpile because it's no longer in the map.
• Test seams __testOnlyGetTranspileCallCount() and __testOnlyGetInFlightTranspileCount() let the dedup tests assert "exactly one transpile call serviced N concurrent same-path readers" and "the slot released on settle" directly rather than inferring it from timing.

Same shape as CS-10947 for CachingDefinitionLookup. Stacks well with CS-11028: the inflight dedup catches the "many readers, one transpile" case; the generation discard catches the "transpile completed against pre-invalidate bytes" case.

Linear: https://linear.app/cardstack/issue/CS-11029

Test plan

• CI passes the new `Realm.#moduleCache in-flight transpile dedup` module in `module-cache-race-test.ts`:
  • N concurrent same-path readers trigger exactly one transpile call
  • concurrent different-path readers each trigger their own transpile (no false coalesce)
  • in-flight entry survives an unrelated path's invalidate
  • in-flight entry is dropped when its own path is invalidated; later caller starts a fresh transpile
  • identity-checked cleanup: A in-flight + invalidate + B in-flight + A settles → B survives
  • errored transpile is shared with concurrent waiters and the in-flight slot releases on rejection
• Existing CS-11028 race tests still pass.
• Existing module-cache regression tests in `realm-endpoints-test.ts` still pass (etag, 304, content-type).

🤖 Generated with Claude Code

[github-actions]

Host Test Results

    1 files      1 suites   1h 46m 37s ⏱️
2 644 tests 2 599 ✅ 15 💤  0 ❌ 30 🔥
2 663 runs  2 588 ✅ 15 💤 30 ❌ 30 🔥

Results for commit 5e2aa2b.

For more details on these errors, see this check.

Realm Server Test Results

    1 files      1 suites   14m 55s ⏱️
1 307 tests 1 289 ✅ 0 💤 18 ❌
1 386 runs  1 368 ✅ 0 💤 18 ❌

Results for commit 5e2aa2b.

For more details on these errors, see this check.