feat(llm): add image/audio output caps, ImageGenOptions, public OneChunkStream

[lIang70]

Summary

Extends sdk/llm with the primitives needed by upcoming
image-generation adapters in sdkx (MiniMax / Seedream /
Qwen-Image), without coupling sdk/llm to any specific provider.

• New capabilities CapImageOutput / CapAudioOutput so output
  modalities are declarable in the blacklist-based ModelCaps
  matrix. Existing chat-only catalogs will get these disabled in a
  follow-up sdkx PR.
• ImageGenOptions + WithImageGen() carry size / aspect ratio
  / N / seed / response-format knobs through
  llm.GenerateOptions.
• NewOneChunkStream extracts the previously-private
  oneChunkStream helper from with_caps.go so synchronous-only
  adapters can satisfy llm.StreamMessage with a single shared
  implementation. with_caps.go now delegates to the public
  helper.

Why this PR exists

Three image-generation providers are about to land in sdkx. They
all need:

1. A way to declare image-output capability without inventing a
   parallel sdk/image interface.
2. A typed channel for image-specific knobs (size etc.) that the
   text-oriented LLM interface lacks.
3. A canonical wrapper to expose synchronous image APIs as a
   single-chunk stream so llm.StreamMessage callers don't break.

Keeping these in sdk/llm (rather than copy-pasting per adapter)
makes the matrix consistent and unblocks sdkx.

Test plan

• go test ./sdk/llm/... -count=1 -race — covers the new
  onechunk_test.go (text, tool calls, multimodal parts) and
  the existing with_caps_test.go which now exercises the
  public helper via the streaming-downgrade path.
• go vet ./sdk/...

Follow-ups

• sdkx: catalog updates (disable new caps on chat models) + the
  three image adapters that consume ImageGenOptions and
  NewOneChunkStream.
• tests/conformance: live API integration tests for the image
  providers (gated on env vars).

Made with Cursor