Suppress benign TSan race in SlabAllocator::getRandomAlloc()

Summary:
Add a TSan annotation to the allocSize read in getRandomAlloc().
This sampling function is best-effort: slab headers can transition
between allocation classes during slab release/rebalance, but callers
validate the sampled item before using it. A racing allocSize read
produces at worst an invalid sample (returned as nullptr or rejected
by the caller's lookup validation). The same annotation pattern is
already used in this file for nearby speculative reads.

___

overriding_review_checks_triggers_an_audit_and_retroactive_review
Oncall Short Name: oil

Differential Revision: D103780529

fbshipit-source-id: 106a8ee1830e23aee4c18e00e49c684ebcbecaa0

Author: ts4z

cachelib/allocator/memory/SlabAllocator.cpp

@@ -472,6 +472,12 @@ std::tuple<uint32_t, const void*> SlabAllocator::getRandomAlloc()
   XDCHECK_GE(reinterpret_cast<uintptr_t>(memory),
              reinterpret_cast<uintptr_t>(slab));
 
+  // getRandomAlloc() is best-effort sampling. Slab headers can concurrently
+  // transition between allocation classes while slab release/rebalance is in
+  // progress; callers validate the sampled item before using it. A racing
+  // allocSize read may at worst produce an invalid sample, which is reported as
+  // nullptr below or rejected by the caller's lookup validation.
+  folly::annotate_ignore_thread_sanitizer_guard g(__FILE__, __LINE__);
   const auto allocSize = header->allocSize;
   if (allocSize == 0) {
     return std::make_tuple(0, nullptr);