Coordinator: use render mm_metadata on prefill instead of full kwargs_data

[zhouyou9505]

What would you like to be added

After vLLM #54659, render returns features.mm_metadata (metadata-only MultiModalKwargsItem blobs: placeholder_metadata_fields / keep_on_cpu, without encoder tensors such as pixel_values). Prefill may omit kwargs_data and send mm_metadata when ec_transfer_params is set.

Update the coordinator to:

1. Parse optional mm_metadata from the render response into each MultimodalEntry.
2. Keep encode on full kwargs_data (ViT still needs pixel tensors).
3. On prefill, when ec_transfer_params is non-empty and every entry has MMMetadata, emit mm_metadata and omit kwargs_data; otherwise keep sending kwargs_data for older renderers and non-EC paths.

Encode  = kwargs_data
Prefill = mm_metadata + ec_transfer_params
no EC / old renderers = kwargs_data unchanged

Documented as the follow-up to the vLLM half in docs/communication.md (Optimization: avoid sending pixel data to prefill).

Why is this needed

Prefill only needs lightweight multimodal metadata (for example image_grid_thw for mRoPE) plus EC transfer params. Forwarding full kwargs_data re-ships ~1 MB/image of pixel_values that encode already consumed. Dropping those blobs cuts coordinator-to-prefill payload size on E/P/D multimodal traffic.

Success criteria

• Render stores mm_metadata when present; absent field remains backward compatible.
• Prefill with NIXL EC + metadata uses mm_metadata and omits kwargs_data.
• Prefill without EC (or without metadata) still sends kwargs_data.
• Encode unchanged.
• Unit tests cover the prefer-metadata and fallback paths.
• docs/communication.md describes the mm_metadata contract (not the earlier plain image_grid_thw sketch).