ggml-cpu : use template for argsort

[slaren]

This can make a significant difference for large arrays.

  Device description: 13th Gen Intel(R) Core(TM) i9-13900K
  Device memory: 175085 MB (175085 MB free)

PR:
  ARGSORT(type=f32,ne=[65000,16,1,1],order=0):                  1033 runs -  4128.78 us/run -     8125 kB/run -    1.88 GB/s

master:
  ARGSORT(type=f32,ne=[65000,16,1,1],order=0):                  1033 runs -  7251.82 us/run -     8125 kB/run -    1.07 GB/s