[Autotuner] Add long-lived benchmark worker pool

[choijon5]

Stacked PRs:

• [Autotuner] Add pool benchmark subprocess mode #2360
• [Autotuner] Add process-level pool reuse #2297
• [Autotuner] Confirm suspicious pool rebenchmark timings #2296
• [Autotuner] Rebenchmark pool configs on owner workers #2295
• [Autotuner] Add isolated worker benchmarking #2294
• [Autotuner] Route search rebenchmark through provider #2293
• [Autotuner] Wire pool mode into benchmark provider #2292
• [Autotuner] Add pool precompile manager #2291
• [Autotuner] Add serializable worker benchmark jobs #2290
• ->[Autotuner] Add long-lived benchmark worker pool #2289

---

Add long-lived benchmark worker pool

splitting up this #2128 so that it's reviewable. This is extending the idea used for long lived processed during benchmarking (#2111) to precompilation. Each worker owns its own CUDA context and can be timed out, killed, and respawned independently, so CUDA sticky errors or hung jobs are contained to the worker instead of poisoning the autotune parent process.
Local benchmarking on H100 (run 3 times each) shows that the new mode, "pool", is faster than fork and on par in terms of perf.

Variance in perf and compile time is also slightly lower than fork:

This PR adds the low-level primitive needed by later PRs to parallelize autotune precompile/benchmark work while keeping CUDA failures isolated to worker processes.