feat(memory-saver): flush PyTorch alloc cache and IPC handles after pause

[qywu]

Stacks on #272.

Summary

torch_memory_saver.pause() releases the physical pages backing tensors in saver.region(), but PyTorch's caching allocator still holds onto its own free pool and any NCCL/IPC handles outside that region. From the driver's perspective those pages are still ours, so co-tenants can't use the headroom pause() just freed.

This adds a torch.cuda.empty_cache() + torch.cuda.ipc_collect() immediately after pause() in the scheduler-side handler, so the bytes are actually surrendered to the driver pool.

Verification on H100

Combined with #272 alone the synthetic measurement was 90.7 % reclaim of engine footprint. With this PR + #274 stacked on #272, the full-stack reclaim on a real engine (Qwen2-1.5B, 40,804 MiB footprint, --enforce-eager) is 97.7 % — see #272 description for the per-phase table.

Isolated contribution of this PR can't be precisely separated from #274 in the stacked test (both ran together). What can be said:

• Without empty_cache() after pause(), the PyTorch caching allocator continues to hold the freed blocks. On a steady-state engine that's typically a few hundred MiB; right after a burst that ended in free → realloc, it can be > 1 GiB.
• ipc_collect() adds a small but non-zero amount (NCCL collectives, P2P buffers).
• Release latency stays at 23 ms — empty_cache() is essentially free at pause time.

Test plan

• Capture nvidia-smi memory.used delta with the empty_cache call across /release_memory_occupation — confirmed in the H100 E2E run.
• Verify no perf regression on /resume_memory_occupation (allocator will rebuild the cache lazily on next forward) — resume latency stayed at 18 ms.
• Single-process A/B test isolating just empty_cache's contribution (low priority; feat(memory-saver): wrap CUDA graphs and attention workspaces in saver.region() #274 is the bigger win in the stack).