Make sample recycling synchronization visible to ThreadSanitizer

[cboulay]

ThreadSanitizer reports a timestamp race when a sample is recycled after another thread releases its reference. The existing release decrement followed by an acquire fence is valid C++ synchronization, but TSan does not model the fence. Replace that pair with an acquire-release decrement so the final owner acquires earlier owners’ releases before publishing the sample for reuse.

Fixes #305. This addresses a sanitizer false positive; the investigation did not demonstrate premature sample reuse or corrupted timestamps. LLVM tracks the fence limitation in llvm/llvm-project#52942.

Changes

• Use memory_order_acq_rel in intrusive_ptr_release() instead of a release decrement plus conditional acquire fence. No suppressions or sanitizer-specific annotations are added. This acquires on every decrement, rather than only on final release; performance has not been benchmarked.
• Add an isolated recycling regression: a reader releases one reference, the last owner releases and reuses the same sample, and only then joins the reader. The scheduling flag deliberately uses relaxed ordering so it cannot hide the reference-count synchronization being tested.
• Fix an existing concurrent send-buffer stress test that called factory::new_sample() concurrently despite its single-allocator contract. Serialize only allocation; buffer pushes remain concurrent. This is a separate test-only commit.

Validation

Native macOS arm64, Xcode 27 / Apple Clang 21, Debug, with TSan instrumentation and no suppressions:

Test	Before	After
Isolated recycling test, 10 process runs	10/10 runs warned	0/10 runs warned
Disconnect-during-push regression, 100 process runs, both transport modes	14/100 runs warned	0/100 runs warned

All functional assertions passed in both comparisons. Before-fix warning runs exited unsuccessfully under TSan; after-fix runs passed without warnings.

• Focused [outlet],[open],[reopen],[sync] suite: 1,741 assertions across 20 cases, passed normally and under TSan.
• Full internal suite: 1,069 assertions across 32 cases, passed normally; after correcting the existing stress-test allocator misuse, also passed under TSan without warnings.
• Final [send_buffer],[sample] tests: passed normally and in 10/10 additional TSan runs.
• An initial broad TSan run hit a socket-test Address already in use failure. A retry exposed the existing concurrent-allocation misuse corrected here; the final full internal run passed.
• git diff --check passed.

Windows/Linux validation remains for CI.