[NVBug 5969206][fix] Force-terminate in-flight KV transfers on abort

[nv-yna]

Summary

Fixes a permanent hang in disaggregated serving where cancelled requests with in-flight KV cache transfers leak blocks until the pool is exhausted.

Root cause: When abort arrives for a request in state 21 (kDISAGG_CONTEXT_TRANS_IN_PROGRESS), the C++ cancel_request() only succeeds for queued transfers, not in-flight ones. Failed cancels were retried every iteration indefinitely — the request stayed in transmission state forever, holding KV blocks that were never freed.

Fix: When cancel_request() returns False for an in-transmission request, force-transition it to the completed state (DISAGG_CONTEXT_COMPLETE / DISAGG_GENERATION_TRANS_COMPLETE) and clean up its resources immediately. The orphaned NIXL transfer is discarded by the peer.

This makes abort work unconditionally regardless of the request's state in the KV transfer pipeline, without requiring kv_transfer_timeout_ms to be configured.

Why this is the right fix

We investigated several approaches and found that any fix relying on downstream signals (first token, transfer completion, Phase 2 timeout cleanup) fails due to circular dependencies — once the KV pool is exhausted, the signal needed to trigger cleanup requires free blocks that don't exist.

The only correct approach is to make abort unconditional: when cancellation is requested, the request MUST be terminated regardless of its internal state.

Approach	Works?	Why
kv_transfer_timeout_ms	Partial	Phase 2 cleanup only runs in _send_kv_async which needs new requests → circular dep
Move Phase 2 to main loop	Partial	Still requires timeout to be configured
First-token guard (delay abort)	No	Held requests consume blocks, preventing first token → circular dep
Force-terminate on cancel (this PR)	Yes	No dependencies — abort is immediate and unconditional

Changes

• _force_cancel_in_flight_request(): New method that force-transitions state 21→22 (or 9→12) and cleans up resources
• _handle_canceled_requests(): When _try_cancel_request() fails, calls force-cancel instead of retrying. Removes the still_pending_canceled_ids retry loop entirely.

Reproduction

Same as #12313. 1P1D disagg with Qwen/Qwen3-0.6B on L40 PCIe, 4+ cancel cycles with concurrency=300.

Test plan

• Verify with L40 PCIe reproducer: 4 cancel cycles + probe should pass
• Verify with Dynamo + ISL=8000 reproducer (which failed with all other fixes)
• Verify no regression on normal disagg serving benchmark
• Verify orphaned NIXL transfers don't cause peer-side issues

🔗 Full analysis: NVBugs 5969206
🔗 Related: #12313 (Phase 2 circular dependency fix — complementary but insufficient alone)