Keep track on the latest CUDA op impl

[xytintel]

We should keep up with the latest CUDA optimizations.

• aten::gather: faster gather implementation pytorch/pytorch#151490 Update scatter gather and index implementation #1643
• aten::gather: Support more dtypes for input, indices in gather pytorch/pytorch#151822 Update scatter gather and index implementation #1643
• Loops kernels: [ATen][CUDA] Implement 128 bit vectorization v2 pytorch/pytorch#145746
• h2d: Tensor .cuda() very slow with specific array sizes pytorch/pytorch#153176
• Add _foreach_fill_ ops pytorch/pytorch#150092
• cat: [aten] 8 bytes aligned vector loads for bf16 and fp16 dtypes in torch.cat pytorch/pytorch#150233