[Opt] cache thread-local CUDA stream for lock acquisition kernels

[rhdong]

• Replace per-call cudaStreamCreate/Destroy in lock_read(), lock_update(), and lock_update_read() with a thread_local cached stream. Eliminates CUDA driver contention when multiple threads acquire locks concurrently, which caused triple-group mode to underperform R/W lock in mixed workloads (0.93x) despite allowing concurrent updaters.

• CI tests passed on the local machine.

• Triple-Group vs R/W Lock Concurrency Benchmark

  Config: dim=16, capacity=128M, HBM=16GB, λ=0.75, batch=64K, 200 batches/thread, H100 NVL

  With Stream Cache Optimization

  Workload	Threads	TG (B-KV/s)	RW (B-KV/s)	Speedup
  Read-heavy	8F/1U/1I	1.451	1.408	1.03×
  Update-heavy	4F/5U/1I	1.675	0.523	3.21×
  Insert-heavy	4F/2U/4I	0.551	0.459	1.20×
  Update-only	1U	1.054	1.046	1.01×
  Update-only	2U	1.613	1.125	1.43×
  Update-only	5U	2.279	0.591	3.86×
  Update-only	10U	2.569	0.535	4.80×

  Before vs After Stream Cache

  Workload	Before	After	Change
  Read-heavy	1.02×	1.03×	—
  Update-heavy	0.93×	3.21×	fixed
  Insert-heavy	0.98×	1.20×	+22%
  Update-only 1U	1.01×	1.01×	—
  Update-only 2U	1.33×	1.43×	+8%
  Update-only 5U	2.36×	3.86×	+64%
  Update-only 10U	2.14×	4.80×	+124%

[github-actions]

Documentation preview

https://nvidia-merlin.github.io/HierarchicalKV/review/pr-248