refactor generator code

[carlushuang]

need to generalize code generation logic for different direction, precision, arch

• global load/store:

  • support different precision, fp32/fp16(short)/ubyte
  • support 2d/3d load, and have exec mask from different dimension
  • support global_load/buffer_load and accumulate through sgpr/vgpr

• share memory load/store:

  • support 1d/2d load/store from different precision
  • support k pack

• coalescing store:

  • support multiple groups to do coalescing store
  • support fp16/int8 final store out pack operation
  • support some case not need LDS shuffle
  • vector write out support

• mfma main loop:

  • different repeat/step
  • support need inst-schedule or no need inst-schedule
  • support k pack suitable from instruction requirement and precision
  • support share load multiple k_pack at once, then do mfma multiple times
  • pass through LDS

• fma main loop

• thread mapping