Batch PICache Mamba-slot releases

[morluto]

Motivation

PICache currently releases Mamba slots one at a time. Every released slot constructs a separate tensor and calls the allocator independently, producing repeated device work and repeated torch.cat operations.

Before

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

After

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

Modifications

• Collect the Mamba slot IDs released by a cache operation.
• Construct one tensor for the complete collection.
• Return all collected slots through one allocator call.
• Preserve the existing cache-removal and ownership semantics.

Accuracy Tests

Not applicable. The change preserves allocator state and cache contents while batching the release operation.

Speed Tests and Profiling

A focused CPU allocator benchmark was run seven times for each batch size:

Released slots	Before	After	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 state-allocator bookkeeping; they are not end-to-end inference timings.

Test Plan

4 passed

The focused tests passed locally and on a clean DigitalOcean Linux environment at commit 98b7618a06a0aa1089324d49a78d4bc4674d1642.

Coverage verifies:

• Multiple slots are returned in one allocator call.
• Released slot IDs remain correct.
• Empty releases do not invoke the allocator.
• Existing cache-removal behavior is preserved.

Closes #5.