[Feature] Add EPD routing with JSON metadata

[gty111]

Purpose

Add E+PD and E/P/D orchestration through --epd-config, integrating encoder routing into vLLM Router instead of requiring a separate Python EPD proxy.

• Select P/PD before encoding, distribute media across E workers, and forward JSON metadata without PyTorch or tensor deserialization.
• Keep embedding transfers in EC backends; support Mooncake PUSH control and opaque NIXL/example handles. Reuse the existing P/D KV handoff, without forwarding P's EC handles to D.
• Preserve sampling fields and SSE responses; cancel abandoned Mooncake transfers.

Native CLI, static workers, DP=1 initially. Requires vLLM JSON metadata support (vllm-project/vllm#56090). See usage and limitations.

Related to #34. No equivalent open PR found: #103 selects modality-specific prefill pools rather than orchestrating separate encoders.

Validation

Native HTTP/SSE smoke passed for E+PD and E/P/D with mock backends, including duplicate images and encoder failures.

Earlier GPU runs: Qwen3.5-35B-A3B, Mooncake EC RDMA, concurrency 96, CUDA graphs through 16384 (no eager), structured choice; 200 warmup requests plus five repetitions of the first 1,200 MUIRBench questions, with caches retained.

4E + 4PD: existing Python EPD proxy vs. vLLM Router

Metric	Python proxy (historical)	vLLM Router	Change
Mean throughput (req/s)	61.40	93.33	+52.0%
Mean request latency (ms)	1514.0	1019.0	-32.7%
Pooled P99 latency (ms)	5007.1	3051.0	-39.1%
Accuracy, all final answers	58.93%	59.13%	+0.20 pp

Same model, client settings and 8-GPU topology, with the proxy on E's node. The Python baseline was not rerun; different nodes/times prevent attributing the entire speedup to the router. Router had one recovered HTTP 502; latency and accuracy include that final response.

Separate 4E + 4P + 4D validation with NIXL KV achieved 90.79 req/s, 58.98% accuracy, with no request failures. This uses 12 GPUs and is not an equal-resource comparison with 4E + 4PD.

These are subset evaluations, not full MUIRBench scores. GPU runs predate the final code cleanup and have not been rerun afterward.

AI assistance was used for implementation and testing.