vLLM PR adds dynamic MoE expert caching

This is the kind of gritty infra patch that could matter more than flashy model launches if it lands cleanly upstream.

• –The PR is currently open (not merged), so real-world benefit depends on review outcome, kernel compatibility, and production hardening.
• –It directly targets a known vLLM pain point around dynamic MoE offloading, where community discussions previously said this was not supported.
• –The design aligns with the core MoE reality: expert usage is skewed, so an LRU hot-expert cache can cut VRAM pressure with acceptable latency tradeoffs.
• –CPU fallback on miss is a pragmatic choice for single-GPU users, but prefill-heavy or highly diverse routing workloads may still see latency spikes.
• –If follow-up work (MXFP4, disk streaming, two-tier cache, EP/DP integration) ships, this could become a strong unlock for local and budget deployments.