docs(reasoner): clarify raw vLLM video sampling

[sylvesterkaczmarek]

Summary

• document the correct vLLM frame-sampling request shape for video inputs
• show both the OpenAI client extra_body form and equivalent raw /v1/chat/completions JSON
• clarify that fps and do_sample_frames belong inside mm_processor_kwargs, not as standalone top-level chat-completion fields

Why

Issue #190 reports weak physical-plausibility classification while sending fps and do_sample_frames as top-level fields in a raw requests payload. The current Reasoner prompt guide documents video sampling through extra_body={"mm_processor_kwargs": ...} for the OpenAI client, but the vLLM quickstart does not show the equivalent raw JSON shape.

That distinction matters for motion-dependent tasks such as physical plausibility analysis because frame sampling determines which temporal events are presented to the Reasoner. This change makes the raw HTTP configuration explicit next to the vLLM quickstart.

Validation

Documentation-only change. Cross-checked against the current Reasoner prompt guide's mm_processor_kwargs sampling configuration and the request shown in #190.

Addresses #190