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

LMCache in-process connector silently returns wrong KV data for models with fused/packed KV layout (K and V share trailing per-head dimension). No exception or warning. use_gpu_connector_v3=True fails differently with tensor size mismatch. LMCacheMPConnector works as workaround. Demonstrates silent KV cache corruption with layout mismatch.