Optimize tracing memory usage

[bobrik]

There are noticeable amounts of memory allocated in these two:

• workerd::IoContext::makeAsyncTraceScope
• workerd::IoContext::makeUserAsyncTraceScope

Changes here aim to reduce that. In order of appearance:

---

Fix allocation sampling in benchmarks

Disable sampling with interval zero and verify that the thread-local
sampler adopts interval one before collecting allocation statistics.
This prevents benchmarks from reporting zero allocations.

Add a regression test covering repeated sampling activation.
Validated with five test runs.

---

Share one async context frame for request tracing

Install internal and user tracing state together at request entry
points, avoiding a second frame allocation and copy of inherited
storage. Preserve separate tracing keys and existing payload lifetimes.

Add regression tests for inheritance, restoration, and captured frames,
plus a benchmark comparing separate and combined scopes.

• CPU time: 4.83 to 2.38 us/op.
• Native allocation volume: 896 to 448 B/op.
• Native allocations: 6 to 3 per operation.
• Allocation volume also halves with 4 and 16 inherited entries.

---

Reduce allocations for unobserved tracing scopes

Capture the originating IoContext ID for unobserved internal and user
tracing scopes instead of allocating span payloads and ownership nodes.
Preserve context masking, foreign-request fallback, and cached span
identity. Create a no-op user span only when requested.

Add regression tests for ownership, span identity, and overlapping
requests, plus a benchmark comparing legacy and compact scope state.

• CPU time: 6.37 to 6.41 us/op, within run-to-run variation.
• Native allocation volume: 1.12 to 0.92 KiB/op (18.0% less).
• Native allocations: 8.03 to 5.05 per operation (about 3 fewer).

---

Store small async context frames inline

Keep up to two storage entries in the frame object and use the existing
hash table for larger contexts. Preserve inheritance, duplicate-key
updates, and dead-key cleanup.

Update memory tracking and add tests for promotion and GC lifetimes.
Extend allocation and lookup benchmarks around the inline threshold.

• Native allocation volume: 448 to 256 B/op (42.9% less).
• Native allocations: 3 to 1 per operation.
• Two-entry lookup: 4.28 to 2.99 ns/op.

[clanky-the-clanker]

✅ Approved — No issues found. Inline frame storage, the combined request trace scope and the compact unobserved scope state keep the existing behaviour.

_{The decision is advisory: this comment never approves, blocks or requests changes on the pull request.}

---

✅ All checks passed

🔒 Security: ✅ LGTM

No security concerns. Isolation between requests still holds for unobserved scopes: a frame from another request falls back instead of exposing its span.

---

📖 AGENTS.md: ✅ LGTM

No AGENTS.md updates are needed for this change.

---

📋 Code Quality: ✅ LGTM

I checked the production changes in IoContext trace scopes, worker-entrypoint, api/tracing.c++ and AsyncContextFrame, and found no defects.

• AsyncContextFrame inline storage:
  • Inherited entries are copied through upsertStorageEntry.
  • Dead keys are skipped on insert and cleared in both the inline and the table representations.
  • The two inline slots are promoted to the hash table when a third distinct key arrives. A duplicate key updates its value in place.
  • get and the upsert path both match keys by pointer identity, so they agree.
• Unobserved scopes:
  • These scopes now keep only the ID of the IoContext they came from.
  • When that ID is the current context, getCurrentTraceSpan and getCurrentUserTraceSpan return an unobserved span, as the old dummy span did.
  • Frames captured from a different request fall back the same way they did before.
• Request entry points: Every place that built the internal and user scopes separately now calls makeRequestTraceScope. No caller of the old two-step setup is left.

---

⚡ Performance: ✅ LGTM

The changes cut down allocations as described. The linear search over the two inline slots is cheaper than a hash lookup at this size.

---

📝 Documentation: ✅ LGTM

The docs/jsg.md update matches the new storage behaviour.

---

_{Reviewed commit: 882f24d5}

[jasnell]

clanky review