feat(qwen3.5): add multimodal tower and projector training controls

[xsuler]

Motivation

Qwen3.5 dense and MoE multimodal models can run frozen image GSPO, but their adapters do not expose the common configure_multimodal_training() contract. The generic tower/projector CLI options therefore cannot unfreeze the Qwen vision tower and merger or assign them independent LR schedules. Model-side trainability controls and checkpoint verification remain the missing pieces.

Proposed feature

Implement the shared multimodal training contract for Qwen3.5 dense and MoE adapters:

• classify the vision transformer as the tower and the vision merger/projector as the projector group;
• keep both media groups frozen by default;
• support --unfreeze-mm-tower and --unfreeze-mm-projector/--unfreeze-mm-merger;
• apply independent base LR, minimum LR, decay steps, and decay style to each group;
• set requires_grad and train/eval mode consistently;
• include unfrozen media parameters in rollout policy synchronization;
• preserve text, tower, and projector weights in standalone checkpoints.

Acceptance criteria

• Qwen3.5 dense and MoE adapters expose the common multimodal training hook.
• Frozen image GSPO behavior and metrics remain unchanged.
• Unfrozen image GSPO produces nonzero tower/projector gradients with the configured LRs.
• Frozen checkpoints preserve original media tensors; unfrozen checkpoints contain changed media tensors.
• Save/reload/serve round trips preserve the trained tower and projector values.
• Unit tests cover parameter classification, freeze state, optimizer groups, and unsupported/missing media modules.

Optimizer and scheduler state restoration is outside this issue; the required checkpoint round trip covers model weights.

Related work

• General multimodal training controls: Generalize multimodal tower/projector training controls and checkpoint round trips #489
• Gemma 4 multimodal support: Support Gemma 4 multimodal serving and agentic training #478