[Bugfix] Fix memory leak and broken metrics in UCMBlendConnector

Two functional bugs in the blend connector:

1. Memory leak: requests_blend_meta is never cleaned up for finished
   requests. build_connector_meta pops from self.requests_meta but
   not from self.requests_blend_meta, causing unbounded growth
   proportional to the total number of completed requests. Each entry
   holds BlendRequestMeta with block hashes and chunk metadata, so
   this is not trivial in long-running serving scenarios.

2. Broken metrics: self.monitor.update_stats() references an undefined
   attribute. The parent class UCMDirectConnector uses the module-level
   ucmmetrics.update_stats() API, but the blend connector calls
   self.monitor which does not exist, causing AttributeError on any
   blend request with metrics enabled.

Signed-off-by: supermario_leo <leo.stack@outlook.com>

Author: SuperMarioYL

ucm/integration/vllm/blend_connector.py

@@ -18,6 +18,7 @@
     UCMDirectConnector,
 )
 from ucm.logger import init_logger
+from ucm.shared.metrics import ucmmetrics
 from ucm.sparse.blend.blockwise_rope import block_wise_rope_forward
 
 if TYPE_CHECKING:
@@ -381,8 +382,7 @@ def get_num_new_matched_tokens(
             f"chunks cache total hit: {chunk_hit_blocks}, "
         )
         if self.metrics_config:
-            self.monitor.update_stats(
-                "ConnStats",
+            ucmmetrics.update_stats(
                 {"interval_lookup_hit_rates": chunk_hit_blocks / max_blk_num},
             )
 
@@ -462,6 +462,7 @@ def build_connector_meta(
         # clear finished request
         for request_id in scheduler_output.finished_req_ids:
             self.requests_meta.pop(request_id, None)
+            self.requests_blend_meta.pop(request_id, None)
 
         return UCMBlendConnectorMetadata(requests_dispatch_meta)