[Bugfix] Replace assertion with guard in finished_sending KV transfer

[Jackie2049]

Summary

Fixes #1

In disaggregated prefill/decode (PD separation) deployments, a race condition exists in the scheduler's handling of finished_sending from the KV connector output.

Problem

At vllm/v1/core/sched/scheduler.py:2153:

for req_id in kv_connector_output.finished_sending or ():
    assert req_id in self.requests  # ← hard crash on race condition
    self._free_blocks(self.requests[req_id])

When a request is aborted (user cancellation, timeout, preemption, or KV load failure) while a KV transfer is still in flight, the request is removed from self.requests. When the KV connector later reports finished_sending for that request, the assertion fires and crashes the entire serving process.

Timeline of the race

1. Request A starts KV transfer to remote instance
2. Request A is aborted (e.g., client disconnect, timeout)
3. Scheduler removes A from self.requests, frees A's blocks
4. KV connector completes send, reports finished_sending = ["A"]
5. assert req_id in self.requests → AssertionError → crash

Fix

Replace the assertion with a defensive guard:

if req_id not in self.requests:
    logger.warning(
        "Request %s not found when finishing KV send; "
        "may have been aborted during transfer.", req_id)
    continue

This is safe because:

• Aborted requests already have their blocks freed during the abort path (_free_request)
• Skipping the duplicate free is the correct behavior
• The continue ensures remaining requests in the loop are still processed

Testing

• Verified the fix compiles without errors
• Unit test for the race condition can be added if maintainers request