Use deformable_detr kernel from the Hub (#36853)

* Use `deformable_detr` kernel from the Hub

Remove the `deformable_detr` kernel from `kernels/` and use the
pre-built kernel from the Hub instead.

* Add license header

* Add `kernels` as an extra `hub-kernels`

Also add it to `testing`, so that the kernel replacement gets tested
when using CUDA in CI.

utils/check_build.py

@@ -21,8 +21,6 @@ from pathlib import Path
 FILES_TO_FIND = [
     "kernels/rwkv/wkv_cuda.cu",
     "kernels/rwkv/wkv_op.cpp",
-    "kernels/deformable_detr/ms_deform_attn.h",
-    "kernels/deformable_detr/cuda/ms_deform_im2col_cuda.cuh",
     "kernels/falcon_mamba/selective_scan_with_ln_interface.py",
     "kernels/falcon_mamba/__init__.py",
     "kernels/__init__.py",

utils/not_doctested.txt

@@ -475,7 +475,6 @@ src/transformers/models/deberta/modeling_tf_deberta.py
 src/transformers/models/deberta_v2/modeling_tf_deberta_v2.py
 src/transformers/models/decision_transformer/modeling_decision_transformer.py
 src/transformers/models/deformable_detr/convert_deformable_detr_to_pytorch.py
-src/transformers/models/deformable_detr/load_custom.py
 src/transformers/models/deit/convert_deit_timm_to_pytorch.py
 src/transformers/models/deprecated/bort/convert_bort_original_gluonnlp_checkpoint_to_pytorch.py
 src/transformers/models/deprecated/mctct/configuration_mctct.py