Add HF Molmo2 loader and logit parity tests

- molmo2_loader.py: HF Molmo2 → MultimodalTransformer weight converter
  (embedding split, fused QKV/MLP split, patch-embedding permute)
- molmo2_parity_test.py: vision-encoder parity vs HF reference
- molmo2_logits_parity_test.py: embedding + LM + full-pipeline logit parity
  (4B/8B pass; O-7B skipped pending per-layer YaRN support)
- Fix transformers 5.x RoPE compat: re-register ROPE_INIT_FUNCTIONS['default']
  and reinit non-persistent inv_freq buffers after from_pretrained fast-init

Author: jason718

src/olmo_core/nn/vision/__init__.py

@@ -10,6 +10,7 @@
     VisionConnectorConfig,
 )
 from .image_vit import SiglipVisionTransformer, VisionTransformer
+from .molmo2_loader import molmo2_hf_state_dict_to_multimodal_transformer
 from .multimodal import MultimodalTransformer, MultimodalTransformerConfig
 
 __all__ = [
@@ -23,4 +24,5 @@
     "VisionConnector",
     "MultimodalTransformerConfig",
     "MultimodalTransformer",
+    "molmo2_hf_state_dict_to_multimodal_transformer",
 ]