CUDA: add conv_2d_dw

[am17an]

Similar to the vulkan implementation, posting some performance numbers from test_backend_ops

CPU

Device description: AMD Ryzen 7 3800XT 8-Core Processor
Device memory: 15968 MB (15968 MB free) 
CONV_2D_DW(ne_input=[512,512,256,1],ne_kernel=[3,3,1,256],stride=1,padding=1,dilation=1,cwhn=0):                16 runs - 101635.44 us/run -   524297 kB/run -    4.92 GB/s 
CONV_2D_DW(ne_input=[512,512,256,1],ne_kernel=[3,3,1,256],stride=1,padding=1,dilation=1,cwhn=1):                32 runs - 44502.06 us/run -   524297 kB/run -   11.94 GB/s

vs CUDA

  Device description: NVIDIA GeForce RTX 3090
  Device memory: 24575 MB (23306 MB free)

  CONV_2D_DW(ne_input=[512,512,256,1],ne_kernel=[3,3,1,256],stride=1,padding=1,dilation=1,cwhn=0):               576 runs -  1853.15 us/run -   524297 kB/run -  269.82 GB/s
  CONV_2D_DW(ne_input=[512,512,256,1],ne_kernel=[3,3,1,256],stride=1,padding=1,dilation=1,cwhn=1):               512 runs -  2176.38 us/run -   524297 kB/run -  233.33 GB/s

Test pass with test_backed_ops -o CONV_2D_DW, also did some manual tests

[JohannesGaessler]

Not that it was introduced in this PR, but is the logic in ggml_is_contiguous_channels correct? I think the comparisons should be >= rather than > to cover cases where one of the dimensions is 1. @Acly since you are the one who added that function, do you agree?

[Acly]

Not that it was introduced in this PR, but is the logic in ggml_is_contiguous_channels correct?

I believe it does the intended thing, also for typical 1D tensors. I used > so that in the (perhaps unlikely) case of ambiguity when both spatial dimensions are 1, it would be "biased" to report false, and select the non-permuted kernel. Making both comparisons use >= would change that behavior.

// 1D tensor, channels most contiguous in memory
x = ggml_new_tensor(ctx, TYPE_I8, C, W, 1, 1);
x = ggml_permute(ctx, x, 2, 0, 1, 3);   // -> ne = W, 1, C, 1
ggml_is_contiguous_channels(x);         // -> true, nb = C, C*W, 1, C*W

Tensor with C 1 H 1 memory layout will not work though. For that case, changing only the nb1 > nb0 check to use >= instead would enable it, without any downsides that I can see.