[train][multimodal][1/3] Add vision support to generate() in new inference stack

[nithinvc]

Summary

1/3 PRs for for #1493 - multi-turn VLM generator

Adds multimodal generation support to the inference client, enabling RemoteInferenceClient.generate() to forward multi-modal features (image hashes, placeholder ranges, kwargs) from vLLM's render endpoint through to the generation endpoint.

• Thread mm_features through RemoteInferenceClient.generate() and _generate_single(), conditionally attaching them as "features" in the HTTP payload to the vLLM server
• Add mm_processor_cache_gb=0 to vLLM CLI args to disable the multimodal processor cache. Required otherwise vLLM won't return multi-modal features for repeated image rendering (/render is not idempotent).
• Add unit test verifying mm_features are forwarded in the HTTP payload via mock server
• Add GPU integration test (test_generate_with_multimodal_features_red_square) that exercises the full render -> generate round-trip with a VLM

Test plan

• Existing test_remote_inference_client.py tests pass: uv run pytest tests/backends/skyrl_train/inference_servers/test_remote_inference_client.py -v
• New TestMultiModalGeneration test passes: verifies mm_features reach the server payload
• GPU integration test passes (requires local vLLM): SKYRL_LOCAL_VLLM=1 uv run --isolated --extra dev --extra fsdp pytest tests/backends/skyrl_train/gpu/gpu_ci/inference_servers/test_vlm_inference_generation.py -m vllm -v

---

[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.

[SumanthRH]

SKYRL_LOCAL_VLLM=1 uv run --isolated --extra dev --extra fsdp pytest tests/backends/skyrl_train/gpu/gpu_ci/inference_servers/test_vlm_inference_generation.py -m vllm -v

What is SKYRL_LOCAL_VLLM @nithinvc? Tests should pass with vllm 0.19.0

[nithinvc]

A temporary flag till vllm-project/vllm#38405 gets merged in. It's been approved but the auto-merger hasn't merged it in (there's some test timeout unrelated to my changes). The tests will likely only pass with vllm 0.20.0 when this is in

[SumanthRH]

Ok sounds good. let's get this in anyways