[train] VLM SFT support on Megatron backend (Qwen3-VL)

[s-chundi]

What

Adds vision-language-model (VLM) support to the SFT trainer on the Megatron backend, starting with Qwen3-VL. The Megatron path previously supported text-only SFT; the FSDP path already had VLM support. This PR brings the Megatron side up to parity for SFT by reusing the existing VLM data infrastructure (TensorList, Experience.pixel_values/image_grid_thw, replay-buffer image fields).

The vision tower runs on the first pipeline stage; pixel_values / image_grid_thw flow through the same TrainingInputBatch → Experience path the FSDP backend already uses, and are passed into GPTModel.forward as Qwen-style kwargs.

Changes

• skyrl/utils/tok.py — check_is_vlm() and get_processor() helpers.
• skyrl/train/sft_trainer.py — detect VLMs in setup(), build an HF processor, thread it through tokenize_chat_example / _tokenize_chat_last_assistant (now return_dict=True with a small _unbatch helper), emit image tensors as a TensorList in collate_sft_batch, force sequential tokenization for VLMs (the processor does not round-trip through the spawn worker pool), and disable sequence packing / microbatch-padding removal.
• megatron_model_wrapper.py — is_vlm flag, VLM constraint asserts, and pixel_values/image_grid_thw passthrough in both forward and forward_backward_mini_batch.
• megatron_worker.py — VLM detection off the HF config, plumb is_vlm into the wrapper, and forward image tensors through the micro-batch builders (policy forward/train + ref forward).
• tests — test_megatron_vlm_init.py (GPU: Megatron VLM forward vs HF reference, and a 5-step SFTTrainer train loop asserting loss decreases) and VLM cases in test_sft_tokenization.py (CPU: tokenization + collation).

(a) Constraints carried over from the FSDP VLM path

3D RoPE and multimodal token positions mean the following are unsupported and asserted/forced off:

• sample / microbatch packing (remove_microbatch_padding, use_sequence_packing)
• context parallelism (CP)
• sequence parallelism (SP)
• mixed text+image microbatches (batches are assumed homogeneously imaged)

(b) Parity / correctness

test_megatron_vlm_forward checks the Megatron VLM forward against a plain HuggingFace AutoModelForImageTextToText forward over the same tokens + images (finite, aligned log-probs). test_vlm_train runs 5 SFT steps and asserts the loss decreases.

(c) Parallelism combos exercised in tests

• TP=2, PP=1 (forward parity)
• TP=1, PP=2 (5-step train loop)

(d) Not supported

CP, SP, sample/microbatch packing, mixed text+image batches, MoE+VLM.