When sampling triggers, we skip SOO and allocate a backing array. We must do this because the HashtablezInfoHandle is part of the heap allocation (along with the control bytes and slots). By default, we sample 1 in ~1024 hashtables when sampling is enabled. This will impact performance because (a) we won't benefit from SOO so we would have worse data locality (more cache/TLB misses), and (b) the backing array capacity will be 3 instead of 1 so (b.1) we skip the rehash after the second insertion and (b.2) we potentially waste up to two slots worth of memory.

We also add an soo_capacity field to HashtablezInfo to allow for distinguishing which sampled tables may otherwise have been SOO - this will allow us to know approximately what fraction of tables are in SOO mode.

PiperOrigin-RevId: 617252334
Change-Id: Ib48b7a4870bd74ea3ba923ed8f350a3b75dbb7d3