ggml : refactor forward_dup for cpu backend

[ngxson]

Ref discussion: #15783 (comment)

Refactor this code by using cpp template function

I tested this code by running test-backend-ops against CPU <--> Metal/CUDA/Vulkan, but maybe there are some cases missed from the test. Would be nice if you can have a deeper look, thanks!

[ngxson]

I did a pref test between master & PR. While the result fluctuate quite a lot from one run to another, I can see that the peak performance stays the same:

master:
  CPY(type_src=f32,type_dst=f16,ne=[512,3072,1,1],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]):                10021 runs -   101.79 us/run -     9216 kB/run -   86.41 GB/s
  CPY(type_src=f32,type_dst=f32,ne=[8192,512,2,1],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]):                 1290 runs -   878.65 us/run -    65536 kB/run -   71.41 GB/s
  CPY(type_src=f32,type_dst=f32,ne=[3072,512,2,1],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]):                 3762 runs -   298.64 us/run -    24576 kB/run -   78.60 GB/s
  CPY(type_src=f32,type_dst=q4_0,ne=[8192,512,2,1],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]):                 900 runs -  1464.49 us/run -    37376 kB/run -   24.43 GB/s
  CPY(type_src=q4_0,type_dst=f32,ne=[8192,512,2,1],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]):                2025 runs -   527.50 us/run -    37376 kB/run -   67.61 GB/s


PR:
  CPY(type_src=f32,type_dst=f16,ne=[512,3072,1,1],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]):                12754 runs -    93.92 us/run -     9216 kB/run -   93.65 GB/s
  CPY(type_src=f32,type_dst=f32,ne=[8192,512,2,1],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]):                 1806 runs -   570.91 us/run -    65536 kB/run -  109.90 GB/s
  CPY(type_src=f32,type_dst=f32,ne=[3072,512,2,1],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]):                 3420 runs -   303.12 us/run -    24576 kB/run -   77.43 GB/s
  CPY(type_src=f32,type_dst=q4_0,ne=[8192,512,2,1],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]):                 675 runs -  1552.52 us/run -    37376 kB/run -   23.05 GB/s
  CPY(type_src=q4_0,type_dst=f32,ne=[8192,512,2,1],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]):                1350 runs -   743.66 us/run -    37376 kB/run -   47.96 GB/s