LMCache issue #4463: Silent KV cache corruption on vLLM 0.26 fused/packed KV layout

The mechanism behind the class of bug where an engine-side KV layout change silently invalidates every external cache. 'On vLLM 0.26, the in-process connector (LMCacheConnectorV1) silently returns wrong KV data for models whose KV cache uses vLLM's fused/packed ("unified") layout, where K and V share the trailing per-head dimension. A request served from the LMCache CPU tier produces different output than the same request computed from scratch. There is no exception and no warning - generations are simply wrong.' Enabling use_gpu_connector_v3 is not a workaround; it fails differently with a hard tensor-size mismatch on the first store (1024 vs 2048 at dim 3). Corroborated and generalised by ai-dynamo/dynamo#14416, which states the trigger was vLLM PR #44455 changing the layout from [num_blocks, 2, block_size, num_heads, head_size] to [num_blocks, num_heads, block_size, 2*head_size], and that the LMCacheMPConnector has the same class of bug on hybrid models where 'the packed layout bypasses the required KV-group edit'.