Jeonghyun/ckb add remining fwd conv device ops

[JH-Leon-KIM-AMD]

Proposed changes

This PR adds factory support for the remaining forward convolution device operations in CK Builder.

Jira Ticket: https://amd-hub.atlassian.net/jira/software/c/projects/ALMIOPEN/boards/319/backlog?selectedIssue=ALMIOPEN-350

Task 350 - Add remaining forward convolution device operations:

1. DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK

   • Added DL (Direct Load) factory specialization for NHWC layout-specific convolution
   • Created new DL-specific algorithm descriptor with 30 template parameters
   • Added test helper function and 3 test cases covering DEFAULT and FILTER_1X1_PAD0 specializations
   • All tests passing (15/15 builder tests)

2. DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor

   • Added Large_Tensor factory specialization for N-dimension splitting (large-than-memory tensors)
   • Implemented macro collision workaround using pragma push/pop for GridwiseGemmTemplateParameters
   • Reuses existing XDL algorithm descriptor (42 identical template parameters)
   • Added test helper function and 2 test cases covering DEFAULT and FILTER_1X1_PAD0 specializations
   • All tests passing (15/15 builder tests)

This completes Task 350 - all 4 forward convolution device operations are now supported in CK Builder:

• ✅ DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle (by Ville)
• ✅ DeviceGroupedConvFwdMultipleD_Wmma_CShuffle (by Ville)
• ✅ DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK (this PR)
• ✅ DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor (this PR)

Checklist

• [x] I have added tests relevant to the introduced functionality, and the unit tests are passing locally
• [ ] I have added the test to REGRESSION_TESTS list defined at the top of CMakeLists.txt in tests/CMakeLists.txt, IF the test takes more than 30 seconds to run
• [x] I have added inline documentation which enables the maintainers with understanding the motivation
• [ ] I have removed the stale documentation which is no longer relevant after this pull request
• [ ] (If this is user-facing) I have added release notes which provide the end users with a brief summary of the improvement from this pull request
• [x] I have run clang-format on all changed files
• [x] Any dependent changes have been merged

Discussion

Design Decisions

1. DL Algorithm Descriptor (30 parameters)

• DL uses VALU instructions instead of XDL matrix cores, requiring different parameter structure
• Fixed NHWC_KYXC_NHWK layout only (not flexible like XDL)
• Created separate descriptor type: ConvAlgorithm_DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
• New DL-specific concepts: DlThreadConfigDescriptor, DlThreadClusterDescriptor, DlBlockTransferK0M0M1K1Descriptor, etc.

2. Large_Tensor Descriptor Reuse (42 parameters)

• Large_Tensor has identical template parameters to regular XDL CShuffle
• Only difference: internal SplitN=true flag in device operation (not exposed in factory interface)
• Reuses existing descriptor: ConvAlgorithm_DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
• No new descriptor or concepts needed

3. Macro Collision Workaround

• Both device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp (line 41) and device_grouped_conv_fwd_multiple_d_xdl_large_tensor_cshuffle.hpp (line 51) define GridwiseGemmTemplateParameters macro without #undef
• Used #pragma push_macro/#pragma pop_macro to isolate Large_Tensor header's macro scope
• This may need to fix on CK headers