[skyrl][tinker] Multi-modal Tinker Sampling

[nithinvc]

Summary

Adds VLM sampling support to RemoteInferenceClient sample endpoint. Finalizes inference side changes for #1200.

• Extract _render_for_sample to handle both text-only and image-containing prompts. For text-only prompts, it flattens chunk tokens directly. When images are present, it calls /v1/chat/completions/render to process images, then splices placeholder tokens into the pre-tokenized text stream with adjusted offsets.
• Update sample() to pass multi-modal features through to the generate payload when present.

Test plan

• Verify text-only sampling still works (no render call made, features is None)
• Verify image sampling works end-to-end (render call is made, placeholder tokens are spliced correctly, features are included in the generate payload)
• New multi-modal sampling tests in test_vlm_inference_generation.py

---

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no potential bugs to report.

View in Devin Review to see 4 additional findings.