Improve DEIMv2 decoder inputs and segmentation pyramid handling

## Summary - ensure the modular DEIMv2 decoder receives explicit embeddings, builds the attention mask, and propagates `decoder_attention_mask` - align decoder query embedding dtypes, honor optional `decoder_inputs_embeds`, and let the segmentation head derive its FPN pyramid dynamically with validation ## Testing - make fix-copies - pytest tests/models/deimv2/test_modeling_deimv2.py::Deimv2ModelTest::test_model -k '' *(fails: `Deimv2ImageProcessor` is not yet implemented)* ------ https://chatgpt.com/codex/tasks/task_b_68dd4bfbbf348336924cdf7a9ce12b38