Add GroupViT to NormalizedConfigManager; fix NormalizedTextAndVisionConfig attribute resolution

[kratos0718]

Summary

Part of #351. GroupViT already has an ONNX export config but was missing from NormalizedConfigManager.

GroupViT is a dual text/vision encoder — same text_config/vision_config shape Pix2StructNormalizedTextConfig already handles — so registering it should have been a one-line reuse of that existing class. It wasn't, because of a real bug this surfaced:

The bug: NormalizedTextAndVisionConfig.__getattr__ prefixes the raw accessor name (e.g. "num_layers") onto the sub-config path, instead of first resolving it through the NUM_LAYERS-style constant mapping to the actual underlying config attribute name ("num_hidden_layers"). Concretely, .num_layers was doing getattr(text_config, "num_layers") instead of getattr(text_config, "num_hidden_layers").

This happened to work for Pix2Struct only because its text_config defines both num_layers and num_hidden_layers as synonyms — coincidence, not correctness. It silently breaks (raises AttributeError) for any model with this same composite shape whose sub-config only has the standard name, which is every other model that would use this class, including GroupViT, OwlViT, OwlV2 and SigLIP.

The fix: resolve the mapped name (getattr(self, attr_name.upper(), attr_name)) before prefixing it with the sub-config path, so the correct underlying attribute name is looked up on the nested config either way.

• Added "groupvit": Pix2StructNormalizedTextConfig to _conf (alphabetically, between gptj and imagegpt)
• Removed groupvit from the TODO: missing normalized configs comment
• Fixed NormalizedTextAndVisionConfig.__getattr__ in the same file

Test plan

• Added test_groupvit_normalizes_expected_attributes: instantiates a real GroupViTConfig() and asserts hidden_size/num_attention_heads/num_layers/image_size/num_channels all resolve correctly against the nested text_config/vision_config
• Added test_pix2struct_num_layers_unaffected_by_the_groupvit_fix as an explicit regression guard, pinning Pix2Struct's existing normalized values so the fix can't silently regress them
• Ran the full test file locally (python -m unittest) against the actual transformers.GroupViTConfig and transformers.Pix2StructConfig (not mocked) — all 3 tests pass

Follow-up

OwlViT, OwlV2 and SigLIP have the exact same shape and are also fixed by this change — I'll open those as separate PRs once this one's reviewed, since they're independent registrations, not because they need anything different.