tp: prove StringPool read path is concurrent-safe under locking

[LalitMaganti]

Adds two ThreadSanitizer stress tests proving the StringPool read path
is safe for the multi-connection design, where string columns are read
on every query thread:

• ConcurrentReadsNoInterns: 8 threads concurrently Get()/GetId() a
  finalized pool (small, block-spanning and >1MB large strings).
• ConcurrentReadsWhileLockedInterns: with locking enabled, reader
  threads do lock-free Get() of pre-existing ids while writer threads
  continuously intern ~2KB strings, forcing several real 4MB-block
  allocations during the read window (the new-block publication path),
  plus occasional large-string appends. The test asserts the bytes of
  every pre-existing id stay intact and self-validates that block
  growth actually occurred.

No production fix was required: the lock-free Get() is safe against
concurrent interns because |blocks_| is a fixed array that never
reallocates (a new block writes a different slot), block bytes are
written forward-only before the Id is published, and large strings are
fully mutex-guarded. Documents this concurrency contract on
set_locking() — including that concurrent interns from multiple threads
are serialized by the mutex and supported (the prior draft contract
wrongly stated otherwise).

[github-actions]

🎨 Perfetto UI Builds

• 5611c75 (2026-09-04 17:16 UTC): https://storage.googleapis.com/perfetto-ci-artifacts/gh-33898808945-1-ui/ui/index.html
• fabebb2 (2026-09-04 18:00 UTC): https://storage.googleapis.com/perfetto-ci-artifacts/gh-33902798298-1-ui/ui/index.html