[vMaroon]

What type of PR is this?
/kind bug

What this PR does / why we need it:

vllm.omitMMKwargs adds "mm_payload": "none" next to "return_mm_kwargs": false on the render copy.

vllm-project/vllm#61427 replaced the return_mm_kwargs render option with mm_payload and kept return_mm_kwargs as a deprecated alias that vLLM ignores when mm_payload is set. With only return_mm_kwargs on the render copy, a client that sends "mm_payload": "tensors" overrides the option on vLLM main, so the render response carries the processed tensors again. Routing stays correct; render time goes back to the full-tensor cost.

The vLLM v0.32 release branch has return_mm_kwargs only, so the render copy carries both fields: v0.32 reads return_mm_kwargs, and versions with vllm-project/vllm#61427 read mm_payload. Both are spliced after the client's fields, and the later duplicate key wins in vLLM's parser, so client values for either field are overridden. v0.32 logs the unknown mm_payload at debug level only. omitMMKwargs: false sends the render copy without either field, as before.

Which issue(s) this PR fixes:

Part of #1552

Test plan

• The render copy gets both fields before its closing brace, client values for either field are overridden, and a non-object body is rejected
• With prefillOnly, the render copy keeps the output-budget rewrite and carries both fields
• Bodies spliced by omitMMKwargs for chat, completions, Responses and Messages, validated with vLLM's request models from releases/v0.32.0 (7960eeb9de) and main (a97d19d792): v0.32 resolves return_mm_kwargs=false and main resolves mm_payload="none", including when the client sent "return_mm_kwargs": true, "mm_payload": "tensors"; no warnings logged
• make presubmit and make test-unit

Release note (write NONE if no user-facing change):

`vllm.omitMMKwargs` also sends vLLM's `mm_payload` render field, so render responses stay free of multimodal tensors on vLLM versions with `mm_payload` when a client sends its own value.