Fix too stringent check on CUDA "fast copy" (can_be_transposed) condition and extend with one more case

[bssrdf]

The additional condition added by #17332 is too strict and excludes some legit cases, e.g., test_cpy(GGML_TYPE_BF16, GGML_TYPE_BF16, {4352, 1, 9216, 1}, {1, 2, 0, 3}, {0, 0, 0, 0}));. This PR loosened that check a bit and added another permuted case which is also a transpose.

Master

 CPY(type_src=bf16,type_dst=bf16,ne=[4352,1,9216,1],permute_src=[1,2,0,3],permute_dst=[0,0,0,0],_src_transpose=0):                     1935 runs -   567.07 us/run -   156672 kB/run -  264.09 GB/s
  CPY(type_src=f32,type_dst=f32,ne=[4352,1,9216,1],permute_src=[1,2,0,3],permute_dst=[0,0,0,0],_src_transpose=0):               1080 runs -   946.85 us/run -   313344 kB/run -  317.06 GB/s
  CPY(type_src=bf16,type_dst=bf16,ne=[21504,4352,1,1],permute_src=[2,0,1,3],permute_dst=[0,0,0,0],_src_transpose=0):                     828 runs -  1280.62 us/run -   365568 kB/run -  273.72 GB/s

This PR

CPY(type_src=bf16,type_dst=bf16,ne=[4352,1,9216,1],permute_src=[1,2,0,3],permute_dst=[0,0,0,0],_src_transpose=0):                     5590 runs -   185.73 us/run -   156672 kB/run -  806.36 GB/s
  CPY(type_src=f32,type_dst=f32,ne=[4352,1,9216,1],permute_src=[1,2,0,3],permute_dst=[0,0,0,0],_src_transpose=0):               2592 runs -   389.38 us/run -   313344 kB/run -  771.01 GB/s
  CPY(type_src=bf16,type_dst=bf16,ne=[21504,4352,1,1],permute_src=[2,0,1,3],permute_dst=[0,0,0,0],_src_transpose=0):                    2208 runs -   460.16 us/run -   365568 kB/run -  761.74 GB/s