Add DETR to NormalizedConfigManager

[kratos0718]

Summary

Part of #351. DETR already has an ONNX export config but was missing from NormalizedConfigManager, so ORTOptimizer can't run on it yet.

• Added "detr": NormalizedTextConfig to _conf (alphabetically, between deit and dinov2)
• Removed detr from the TODO: missing normalized configs comment above it

Unlike the other recent additions on this issue (ConvNeXT, BEiT, LeViT, etc., all vision models needing NormalizedVisionConfig), DETR needs its transformer encoder stats: hidden_size, num_attention_heads, num_hidden_layers. DetrConfig stores these under d_model, encoder_attention_heads, and encoder_layers, but DetrConfig.attribute_map already aliases all three to the names NormalizedTextConfig looks for, so getattr(config, "hidden_size") etc. resolves directly with no custom with_args(...) mapping needed (unlike e.g. BartLikeNormalizedTextConfig, whose config doesn't have this alias).

Test plan

• Added two tests to tests/utils/test_normalized_config.py: one confirming detr resolves to NormalizedTextConfig, and one instantiating a real DetrConfig() and asserting the normalized hidden_size/num_attention_heads/num_layers match d_model/encoder_attention_heads/encoder_layers
• Ran both locally against the actual transformers.DetrConfig (not mocked) — both pass