Add separate GMMA microbenchmark build target

[JRPan]

Summary

• Filter GMMA directories out of the default GPU_Microbenchmark build so they don't break standard compilation
• Add dedicated gmma / GPU_Microbenchmark_gmma Make targets to compile GMMA benchmarks independently
• Add corresponding clean_gmma target for cleanup

Test plan

• Run make GPU_Microbenchmark and verify GMMA directories are excluded
• Run make GPU_Microbenchmark_gmma and verify only GMMA benchmarks are built
• Verify make clean in GPU_Microbenchmark still works for non-GMMA targets

🤖 Generated with Claude Code

[JRPan]

Review

Changes look good. Here's a summary of what this PR does:

src/cuda/GPU_Microbenchmark/Makefile

• Renames SUB_DIRS to ALL_SUB_DIRS to hold all subdirectories, then splits them:
  • GMMA_DIRS — directories ending in gmma/ (via $(filter %gmma/))
  • SUB_DIRS — everything else (via $(filter-out))
• Adds gmma and clean_gmma targets that mirror the existing all / clean targets but operate only on GMMA directories.

src/Makefile

• Adds a top-level GPU_Microbenchmark_gmma target that invokes the new gmma sub-target and moves the resulting binaries into the output directory.

Notes

• The %gmma/ filter pattern will match any subdirectory whose name ends with gmma. If a non-GMMA benchmark name ever ends in gmma it would be incorrectly filtered — unlikely but worth keeping in mind.
• The existing all and clean targets are unaffected for non-GMMA benchmarks; this is a purely additive change.

LGTM ✅