Eliminate harness measurement distortion in TransferCache benchmarks.

[copybara-service]

Eliminate harness measurement distortion in TransferCache benchmarks.

BM_InsertRange, BM_RemoveRange, BM_RealisticBatchNonBatchMutations, and
BM_RealisticHitRate in transfer_cache_benchmark invoked state.PauseTiming()
and state.ResumeTiming() inside every iteration of the benchmark loop,
spending ~60% of CPU cycles in clock_gettime/vDSO syscall overhead (~1300 ns/iter)
and obscuring the underlying 10-30 ns cache operations. In addition,
FakeTransferCacheEnvironment::Insert and Remove allocated temporary std::vector
instances on the heap in each call.

Amortize setup across batches with state.KeepRunningBatch in BM_InsertRange and
BM_RemoveRange, precompute PRNG decisions in BM_RealisticBatchNonBatchMutations
and BM_RealisticHitRate, and replace dynamic std::vector buffers with stack
arrays in FakeTransferCacheEnvironment.

Also:

• Fix mincore_benchmark invalid address (NULL) and double page size
  multiplication bug in memory offset calculations.
• Add size classes 128 B and 256 B to BM_new_delete_fixed in tcmalloc_benchmark.
• Use sized deallocation MockSpan::Delete in linked_list_benchmark.