feat: add MiniCPM-V 4.6 image-to-text support

[Dayuxiaoshui]

What does this PR do?

This PR enables image-to-text training and serving for MiniCPM-V 4.6 checkpoints through AReno's existing MiniCPM adapter.

The implementation:

• aligns MiniCPM-V processor outputs (pixel_values, target_sizes, and patch grouping metadata) with AReno's multimodal feature schema;
• loads and saves the MiniCPM-V NaViT vision tower and merger/projector weights alongside the language model;
• executes the variable-resolution vision encoder, window merger, and final visual-to-language projection inside the AReno model adapter;
• replaces image placeholder token embeddings with projected image features;
• preserves the hybrid Gated DeltaNet/full-attention paths during training, prefill, eager decode, and CUDA Graph decode;
• adds dense recurrent-cache slot management so hybrid attention state is independent of paged KV-cache block IDs and is cleared before slot reuse;
• supports multimodal SFT, GSPO/GRPO-style rollout training, checkpoint round trips, tensor parallelism, and OpenAI-compatible serving; and
• prevents the Qwen3.5-VL adapter from claiming nested MiniCPM-V text configs.

Related issue

Fixes #137

Parent issue: #132

Type of change

• Bug fix
• New feature
• Breaking change (public API / CLI behavior changes in a non-backward-compatible way)
• Documentation update
• Refactoring
• Performance improvement
• Test coverage improvement

How was it tested?

CPU coverage:

pytest -q tests/test_inference_scheduler_cpu.py tests/test_minicpmv46_multimodal_cpu.py
# 34 passed

pytest -q tests/test_registry_cpu.py tests/test_registry_discovery_cpu.py
# 7 passed

CUDA_VISIBLE_DEVICES= pytest -q tests/test_policy_tensor_sync_cpu.py::test_real_gloo_collectives_reshard_train_tp2_to_rollout_tp1 tests/test_policy_tensor_sync_cpu.py::test_real_gloo_collectives_reshard_train_tp1_to_rollout_tp2
# 2 passed

pytest tests/ -k cpu completed with 450 passing tests. Three socket/Gloo failures from the restricted test environment passed when rerun outside the sandbox (with GPUs hidden for the two CPU policy-sync cases). Two existing Qwen3.5-VL CPU tests still fail on the prerequisite Qwen3.5 image-to-text baseline and are unrelated to the MiniCPM-V changes.

Real-model validation used the ModelScope checkpoint OpenBMB/MiniCPM-V-4.6 on NVIDIA H200 GPUs:

• loaded all 779 checkpoint tensors, including 453 vision-tower tensors and 6 merger tensors;
• matched sampled text/vision checkpoint mappings exactly (max_abs_diff=0);
• matched the Transformers reference projected image embeddings exactly (max_abs_diff=0, mean_abs_diff=0);
• served red and blue image questions through the OpenAI-compatible API and returned Red and Blue;
• generated a 16-token image response through CUDA Graph replay, with runtime logs reporting cuda_graph=True;
• completed two consecutive image SFT steps with finite loss and gradient norms, then saved and reloaded the checkpoint successfully;
• confirmed finite, nonzero gradients for all 453 vision-tower gradient tensors and all 6 merger gradient tensors;
• completed a TP=2 image SFT step; and
• completed one end-to-end image GSPO step covering rollout, reward, multimodal training, and policy role switching.

The installed flash_attn package in the validation environment did not export the required attention functions, so GPU validation used --attn-backend native. This is an environment limitation rather than a MiniCPM-V adapter failure.

Checklist

• The PR title summarizes the contribution.
• Linked the related issue in the description (if any).
• Existing tests pass (pytest tests/ -k cpu).
• New behavior is covered by tests.
• Described the test commands run and any hardware limitations.
• Public API / CLI changes are additive and backward-compatible (see CONTRIBUTING.md).