Unify CUDA and HIP kernel sources via compat.cuh portability layer #127

Job	Run time
CPU (linux-aarch64, 2.10.0) / build	16s
CPU (macos, 2.10.0) / build	21s
CUDA (linux-x64, L40S, 11.8.0) / build	2m 38s
CUDA (linux-x64, T4, 12.8.1) / build	3m 4s
CPU (macos, 2.3.1) / build	25s
CUDA (linux-x64, T4, 13.0.2) / build	2m 37s
CPU (linux-x64, icelake, 2.3.1) / build	21s
CPU (linux-aarch64, 2.5.1) / build	11s
CPU (linux-x64, icelake, 2.10.0) / build	18s
CUDA (linux-x64, T4, 11.8.0) / build	3m 13s
CUDA (windows, T4, 11.8.0, 2.7.1, https://download.pytorch.org/whl/cu118) / build	4m 17s
CUDA (linux-x64, L40S, 13.0.2) / build	2m 51s
CPU (linux-x64, 2.10.0) / build	17s
CPU (linux-x64, 2.3.1) / build	13s
CUDA (linux-x64, L40S, 12.8.1) / build	2m 48s
CPU (linux-x64, 2.3.1) / test	10m 1s
CPU (linux-aarch64, 2.5.1) / test	8m 33s
CPU (linux-x64, 2.10.0) / test	15m 26s
CPU (linux-x64, icelake, 2.10.0) / test	9m 12s
CPU (linux-aarch64, 2.10.0) / test	9m 1s
CPU (linux-x64, icelake, 2.3.1) / test	8m 3s
CPU (macos, 2.10.0) / test	11m 7s
CPU (macos, 2.3.1) / test	13m 55s
CUDA (linux-x64, T4, 13.0.2) / test	8m 36s
CUDA (linux-x64, L40S, 11.8.0) / test	3m 47s
CUDA (linux-x64, L40S, 12.8.1) / test	6m 32s
CUDA (linux-x64, L40S, 13.0.2) / test	6m 1s
CUDA (linux-x64, T4, 12.8.1) / test	9m 26s
CUDA (linux-x64, T4, 11.8.0) / test	5m 46s
CUDA (windows, T4, 11.8.0, 2.7.1, https://download.pytorch.org/whl/cu118) / test	14m 20s
	2h 43m 36s

Provide feedback