[fix][model] preserve video input messages

[octo-patch]

Reason: MiniMax-M3 accepts video input, but video message parts are dropped before model invocation.

• Preserve video URL metadata across API, domain, and Eino conversions.
• Expose video capability metadata and validate video inputs against model abilities.
• Add focused coverage for conversions and request validation.

Checks:

• cd backend && go test ./modules/llm/domain/entity ./modules/llm/application/convertor ./modules/llm/domain/service
• git diff --check