[MODEL] add Mage-VL, Muse Glimmer, OLMo 3, SmolLM3, and DeepSeek V3.2 quantization support

[ZX-ModelCloud]

Summary

Add quantization support for Mage-VL, Muse Glimmer, OLMo 3, SmolLM3, and DeepSeek V3.2.

What Changed

• Add model definitions and registrations for Mage-VL, Muse Glimmer, OLMo 3, and DeepSeek V3.2.
• Register SmolLM3 with its Llama-compatible quantization module tree.
• Add multimodal processor and calibration handling for Mage-VL and Muse Glimmer.
• Handle OLMo 3 sliding/full-attention RoPE during layer replay.
• Cover DeepSeek V3.2 DSA, dense layers, routed/shared experts, and auxiliary checkpoint tensors.
• Update required Transformers and Defuser versions.
• Add targeted model-definition, runtime-layout, replay, and regression tests.

Tests

• I added a new simple/fast unit test for this change, or documented why that is not applicable.
• I ran the new targeted test locally before opening this PR.
• I ran any other directly relevant local tests.

Paste the exact test commands and results here:

pytest -q \
  tests/models/test_mage_vl.py::test_mage_vl_model_type_selects_definition \
  tests/models/test_mage_vl.py::test_mage_vl_definition_matches_qwen3_decoder_contract \
  tests/models/test_mage_vl.py::test_mage_vl_processor_load_enables_remote_code \
  tests/models/test_mage_vl.py::test_mage_vl_uses_qwen_style_image_calibration_dataset \
  tests/models/test_muse_glimmer.py::test_muse_glimmer_model_type_selects_definition \
  tests/models/test_muse_glimmer.py::test_muse_glimmer_module_tree_matches_text_decoder_order \
  tests/models/test_muse_glimmer.py::test_muse_glimmer_replay_rebuilds_layer_specific_mask_and_rope \
  tests/models/test_muse_glimmer.py::test_muse_glimmer_replay_matches_dense_text_forward \
  tests/test_model_test_helpers.py \
  tests/test_olmo3_support.py \
  tests/test_smollm3_support.py \
  tests/test_deepseek_v32_support.py
# 30 passed

ruff check \
  gptqmodel/models/definitions/deepseek_v32.py \
  gptqmodel/models/definitions/mage_vl.py \
  gptqmodel/models/definitions/muse_glimmer.py \
  gptqmodel/models/definitions/olmo3.py \
  tests/models/model_test.py \
  tests/models/ovis/image_to_test_dataset.py \
  tests/models/test_deepseek_v32.py \
  tests/models/test_mage_vl.py \
  tests/models/test_muse_glimmer.py \
  tests/models/test_olmo3.py \
  tests/models/test_smollm3.py \
  tests/test_deepseek_v32_support.py \
  tests/test_model_test_helpers.py \
  tests/test_olmo3_support.py \
  tests/test_smollm3_support.py
# All checks passed!

pytest --collect-only -q \
  tests/models/test_mage_vl.py \
  tests/models/test_muse_glimmer.py \
  tests/models/test_olmo3.py \
  tests/models/test_smollm3.py \
  tests/models/test_deepseek_v32.py \
  tests/test_deepseek_v32_support.py \
  tests/test_olmo3_support.py \
  tests/test_smollm3_support.py
# 26 tests collected

Review Requirements

AI-assisted code is welcome.

Every changed file must still be properly reviewed by a human before the PR is opened as ready for review.

• I personally reviewed every file in this diff.
• I checked that the code matches existing project structure, APIs, and conventions.
• I avoided unnecessary monkeypatching and used the project's normal extension points where possible.

Notes

Real-checkpoint quantization and evaluation tests were not executed in the current no-CUDA environment; their test modules collect successfully.