Add GPU kernel backend to optimize rendering performance

[zongwave]

1. add abstract class "Backend" to support different GPU accelerators, CUDA or DPCPP.

2. import inspect module to load CUDA/DPCPP kernel function defined in backends.

3. add ray_sampling CUDA kernel function, tested training, rendering & evaluation 360_v2 image set on NV3070.

   • rendering performance increased about 17% compared with original pytorch ray_sampling.
   • evaluation average quality keep constantly PSNR:26.922/SSIM:0.8233 V.S. PSNR:26.921/SSIM:0.8233 compared with original pytorch ray_sampling
   • training 360_v2/bicycle stop at PSNR: 23.78, the time spent 2.355hours V.S. 2.363hours

4. add cast_ray and grid_encoder DPCPP kernel functions.