[Performance] Batch PICache Mamba-slot releases

[morluto]

Current limitation

PICache releases Mamba state slots one at a time. Each slot constructs a separate device tensor and calls the allocator independently, causing repeated kernel launches and repeated torch.cat work.

slot 1 --> tensor([1]) --> free() --> torch.cat
slot 2 --> tensor([2]) --> free() --> torch.cat
  ...                                  ...
slot N --> tensor([N]) --> free() --> torch.cat

The slot IDs are already available together during cache removal, so this per-slot device work is avoidable.

Desired outcome

Collect all released Mamba slot IDs, construct one device tensor, and return them through one allocator call while preserving cache-removal and ownership semantics.

[slot 1, slot 2, ..., slot N]
              |
              v
       one device tensor
              |
              v
       one free()/torch.cat

Performance evidence

A focused CPU allocator benchmark, repeated seven times per batch size, measured:

Released slots	Per-slot release	Batched release	Improvement
128	3.223 ms	0.599 ms	5.4x
512	8.065 ms	1.186 ms	6.8x
1,024	17.973 ms	2.406 ms	7.5x

These measurements isolate allocator bookkeeping rather than end-to-end inference.