vLLM enables node-shared Engram tables by default and degrades instead of failing startup.

PR #57651 resolves dp_shared_memory automatically when CPU offload is on with data_parallel_size > 1 and elastic EP off, where it previously had to be requested by hand and hard-failed when requirements weren't met. The number making the fallback necessary: DeepSeek-V4.1-Flash's full Engram tables are 188.8 GiB against Docker's default 64 MiB /dev/shm. It compares the tmpfs total, which is identical on every rank so no collective is needed, and degrades to DP head sharding with a warning pointing at --shm-size.