[SkyRL][train] Support prompt_logprobs in /sample in the new inference stack

[nithinvc]

Summary

This PR addresses #1416 and adds prompt log probabilities to the new /sample endpoint in the new remote inference client.

• Add prompt_logprobs and topk_prompt_logprobs support to RemoteInferenceClient.sample(), wiring through to vLLM's /inference/v1/generate endpoint
• Accepts include_prompt_logprobs (bool) and topk_prompt_logprobs (int) from the request body, with fallback to prompt_logprobs for Tinker SampleRequest compatibility

Test plan

• CPU: test_sample — verifies include_prompt_logprobs=True returns per-token logprobs with correct values
• CPU: test_sample_n2 — verifies prompt logprobs work with n=2
• CPU: test_sample_topk_prompt_logprobs — verifies topk_prompt_logprobs=2 returns both fields with correct (token_id, logprob) values
• GPU: test_client_sample — verifies vLLM returns prompt logprobs (non-positive floats)
• GPU: test_client_sample_multiple — verifies prompt logprobs with n=3
• GPU: test_client_sample_topk_prompt_logprobs — verifies that the vLLM top-k tuples contain the actual prompt token

---

[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.