perf(precompute): parallelize independent series updates

[zzylol]

Closes #593.

Splits synchronization into a generation barrier plus per-series ownership:

• established series take the window generation read lock and their own entry lock;
• independent series mutate concurrently;
• same-series updates remain serialized;
• new admission still takes the structural write lock;
• rotation/drain take the write lock and therefore wait for all in-flight entry mutations before detaching the generation;
• eviction-index updates use a narrow auxiliary lock only when EvictOldest is active.

Adds a RunParallel established-series benchmark. Review-host result: ~138 ns/op, zero allocs/op at 64-way benchmark parallelism. go test -race . passes.

Stacked on #601.