Simplify kernel argument of device operator Device(Batched)GemmXdl<>

[poyenc]

I had tried to reduce kernel arguments for following device operators

• DeviceGemmXdl<> (create descriptors on device side)
• DeviceBatchedGemmXdl<> (create descriptors on device side)
• DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<>
• DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<>
• DeviceConvNdBwdDataNwcKxcNwk_Xdl<>

In general, the stateless elementwise operation object & Block2CTileMap are totally redundant parameters for kernels. Thus we can remove them safely in most of the device operators. I only tried to update GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3<> related device operators for now.

To achieve better code reusability, I had decoupled the dependency between BlockToCTileMap_M00_N0_M01Adapt<> and its third type argument: CGridDesc_M_N. Which means we no-longer need to pass CGridDesc_M_N in order to create tile mapping objects.

I also splitted the GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3<> template into 2 templates (also corresponding entry kernels). They are dedicated for different type of device operators.

• GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3<>: For Convolution device operators. No default descriptor creators. Users have to create descriptors by their own.
• GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3_ext<>: For GEMM device operators. Has default descriptor creators for A/B/C matrics. Users have to provide extra A/B/CLayout and GemmSpecialization template arguments.

I found that we may make code more unmanageable if moving descriptor creation logics into device functions. I had done this for DeviceConvNdBwdDataNwcKxcNwk_Xdl<> on branch feature/simplify-karg-for-device-gemm-xdl. But the code is hard to read (the Argument class becomes tediously long) and exists unknown correctness issue since revision: d4efc6a (I think it will take more time to troubleshoot).

This patch requires changes from following pull requests:

• Simplify kernel argument of device operator DeviceGemm_Xdl_CShuffle<> #696
• Add class type support for __builtin_amdgcn_readfirstlane() #711

[poyenc]

Looks like we have correctness issue with DeviceGemm_Xdl_CShuffle<> since revision: 5710567 (should pass the tests in #696). But I didn't touch the DeviceGemm_Xdl_CShuffle<> in this PR.
Shall update the underlying branch of #696 and run CI again.

[poyenc]

I Just fixed the CI server test failure issue in #696