Add PegaFlow external KV cache blog post

[Alex-yang00]

Summary

This PR adds a new blog post for the vLLM x Novita AI collaboration on PegaFlow, an external KV cache service for production LLM serving.

The post covers:

• Why moving KV cache lifetime out of the inference process improves restart behavior and failure isolation
• How PegaFlow pools KV cache across local instances, TP ranks, and remote nodes
• Benchmark results for startup time, local cache sharing, MLA KV deduplication, and RDMA remote reads
• The three-level cache hierarchy with pinned DRAM, remote RDMA-accessible DRAM, and SSD
• vLLM integration through the external KV connector interface
• Quick-start commands and a public reference benchmark from the PegaFlow repository

Assets

Adds figures for:

• PegaFlow architecture
• Startup time comparison
• Rust/Python tail-latency comparison
• Local sharing result summary
• Cross-node RDMA throughput
• Cache-policy comparison

Review notes

The public PegaFlow install commands, connector configuration, P2P flags, and reference benchmark were checked against the novitalabs/pegaflow README and docs. The internal production benchmark numbers should still be confirmed by the Novita AI team before marking this PR ready for merge.

[esmeetu]

We can specify the vLLM version used.

[Alex-yang00]

Thanks! Added a note in the quick-start section that the examples in this post use vllm>=0.20.0.

[esmeetu]

LGTM! There's small suggestion, please take a look.