[feat] add LLaVA-OneVision2 (8B-Instruct) training support

[kcz358]

Motivation

Add training support for the LMMs-Lab LLaVA-OneVision2 8B-Instruct
checkpoint. OV2 ships its modeling and processor code via
`auto_map` (trust_remote_code) and pairs a custom OneVision vision
encoder with a stock Qwen3-8B language model, so most of the
machinery (liger, rmpad, fused LCE) can be reused from the existing
qwen3 plumbing — the OV2-specific work is mostly glue.

Modifications

Model loading

• `mapping_func.create_model_from_pretrained` now forwards
  `trust_remote_code` to `AutoConfig.from_pretrained` and falls
  back to the AutoModelFor* class declared in the checkpoint's
  `auto_map` when the config class is not registered in any HF
  model mapping (the case for all trust_remote_code checkpoints).
• `runner._build_model` passes `extra_kwargs.trust_remote_code`
  through.

Monkey patch infra

• `MonkeyPatcher.apply_monkey_patch[_to_instance]` now skips
  patch_types that are not registered for a given model_type instead
  of raising `KeyError`. This lets us auto-append new patch_types
  from the runner without breaking existing models.

OV2 monkey patches (instance-only, since auto_map)

• New `models/llava_onevision2/` with two patch_types:
  • `liger`: RoPE, RMSNorm, SwiGLU on inner Qwen3 layers; LayerNorm
    on OV2 vision encoder; binds OV2 `causal_lm_forward` with
    `loss_fn="lce"`, `use_rmpad=False`.
  • `rmpad`: class-level qwen3 attention/decoder/model_forward
    rmpad patches, OV2 outer `model_forward` binding, and a rebind
    of `causal_lm_forward` with `use_rmpad=True` that detects the
    liger-bound `loss_fn` (if any) to preserve fused LCE.
• Runner auto-appends `rmpad` patch_type when
  `trainer_args.use_rmpad=True`.
• Stacking matrix:

  use_liger_kernel	use_rmpad	Behaviour
  ✅	✅	rmpad + fused LCE (historical default)
  ✅	❌	fused LCE, no unpad
  ❌	✅	unpad + standard CE
  ❌	❌	stock HF forward

Shared LM loss helper

• New `models/common_ops/loss.py::compute_lm_loss` factors out the
  reusable next-token loss path (LCE / CE, optional rmpad seq-wise
  shifting, Ulysses SP gather). OV2 uses it; future RFCs that split
  liger/rmpad on other models can reuse it too.

Data processor

• New `datasets/processor/llava_onevision2_processor.py` inheriting
  `Qwen3_VLDataProcessor`. Loads the OV2 `AutoProcessor` with
  `trust_remote_code=True`, rewrites
  `<vision_start><video_pad><vision_end>` into per-frame
  `<X.X seconds><vision_start><image_pad>*n<vision_end>` blocks,
  aliases video tensors into the image path (expands
  `video_grid_thw[T,H,W]` into T rows of `[1,H,W]`), builds
  block-layout `patch_positions`, and normalizes `qwen_vl_utils`
  CHW float frames to HWC uint8 so OV2's video processor's
  `PIL.Image.fromarray` branch works.

Misc

• `models/utils.py`: route `llava_onevision2` flops through the
  qwen2 estimator using `config.text_config` (its LM is Qwen3).
• New `examples/llava_onevision2/{example.yaml,run.sh}` end-to-end
  training scaffold.
• New `docs/models/llava_onevision2.md` describing the patch
  composition + data processor; registered in
  `docs/models/index.rst` (also picks up the existing-but-unlinked
  `llava_onevision1_5.md`).

Testing

• Manual run on a single-node 4-GPU box: torchrun + FSDP2 + rmpad +
  liger + sequence packing kicks off training successfully against
  `LLaVA-Video-178K` parquet shards; loss decreases over a short
  smoke run with no NaNs.
• All pre-commit hooks (black + isort) pass.

Commit log

• `feat(model_loading): forward trust_remote_code for auto_map checkpoints`
• `feat(monkey_patch): skip unregistered patch_types gracefully`
• `feat: add LLaVA-OneVision2 model + data processor`
• `docs: add LLaVA-OneVision2 example config and model guide`

Checklist

• Follow commit message convention
• Run `pre-commit run --all-files` and ensure all checks pass
• Format with `black` (line-length=120) and `isort`
• Add unit tests for new functionality (manual smoke test only for now)
• Update documentation