[Feature] Support image-text multimodal input for training

Motivation

Training in AngelSpec is text-only today, so vision-language targets (Qwen3.5,
Qwen3-VL, Hy3-VL) can't be used. We'd like to train DFlash/DSpark drafters for
VLM targets on image-text multimodal input.

Some of the plumbing already exists (media extraction in the dataset, an
image_placeholder in templates, enable_multimodal in the inference config,
multimodal_inputs passed to vLLM). The gap is on the draft-training side,
where block-parallel drafters have no mRoPE support.

Proposed Solution

Support image-text records end to end: accept images in the training data, carry
the target's mRoPE positions through to the draft training forward, and make the
block-parallel draft RoPE mRoPE-aware.

The drafter core shouldn't need to change — KV injection, anchor sampling, and
the losses are all modality-agnostic — and since DFlash, DSpark, and DFly share
a backbone, one implementation should cover all three.

Happy to help implement or test.

Start by tracing media extraction in the dataset, the image_placeholder in templates, and enable_multimodal in the inference config. Follow multimodal_inputs into vLLM, then inspect the draft-training forward and block-parallel draft RoPE path for carrying the target's mRoPE positions. Done means image-text records work end to end for the shared DFlash, DSpark, and DFly backbone.