fix: cancel KVTask immediately when match/put condition not met to prevent memory leak

[xianweihuihuan]

Problem

In _get_match() and the put path, flexkv_manager.get_match() /
get_put() unconditionally creates a KVTask stored in
KVTaskManager.tasks. When the match/put condition is not met
(num_new_matched_tokens == 0 or num_unmatched_tokens == 0), the
task is neither added to req_id_to_task_dict nor cancelled,
causing it to leak in self.tasks until ExpiringDict auto-expires
it after 30 minutes.

Under high concurrency with frequent cache misses, each request
produces a leaked KVTask holding token_ids, slot_mapping, and
token_mask arrays, leading to significant memory pressure.

Fix

Call self.flexkv_manager.cancel(task_ids=[task_id]) immediately
in the early-return branches where the task would otherwise be
abandoned, ensuring prompt cleanup of task resources.

[xianweihuihuan]

Updated: the get path (_get_match) already handles this case
in the original code. This PR only adds the missing cancel call
in the put path (_put_match) when num_unmatched_tokens == 0.