[CUDA][onnx-frontend]add deformconv2d converter

[AIYoungcino]

When compiling and running inference for the Bevformer network using TVM, I encountered an error stating that DeformConv2d is not supported. Although I have added support for DeformConv2d, I am open to other suggestions or recommendations. Please feel free to share your insights.

[AIYoungcino]

with Input: img, Shape: [1, 6, 3, 736, 1280], Data Type: 1 Input: ref_2d, Shape: [1, 22500, 1, 2], Data Type: 1 Input: indexes_3541, Shape: [1, 3541], Data Type: 7 Input: indexes_4256, Shape: [1, 4256], Data Type: 7 Input: indexes_4265, Shape: [1, 4265], Data Type: 7 Input: indexes_5444, Shape: [1, 5444], Data Type: 7 Input: indexes_4043, Shape: [1, 4043], Data Type: 7 Input: indexes_4110, Shape: [1, 4110], Data Type: 7 Input: reference_points_rebatch, Shape: [1, 6, 5444, 4, 2], Data Type: 1 Input: count, Shape: [1, 22500], Data Type: 1 Input: shift, Shape: [1, 2], Data Type: 1 Input: can_bus, Shape: [1, 18], Data Type: 1 Input: prev_bev, Shape: [1, 22500, 256], Data Type: 1 Output: bev_embed, Data Type: 1 Output: all_cls_scores, Data Type: 1 Output: all_bbox_preds, Data Type: 1

Execution time summary on bevformer_small: mean (ms) median (ms) max (ms) min (ms) std (ms) 43680.2000 43684.3969 43686.0125 43670.1906 7.1084

[tqchen]

We are moving toward relax as the new graph IR and relay is mostly in legacy maintaining mode, would be good to see if we can add relax support