experimental: grouped gemm AOT

[saltyminty]

Summary

• Add CuTe DSL AOT export/load helpers and artifact metadata handling.
• Wire CUDNN_FE_AOT_MODE / CUDNN_FE_AOT_DIR into GemmAmax and grouped GEMM wrapper paths.
• Add AOT export/load coverage for GemmAmax and grouped GEMM SM100 APIs, including C++ dlopen/dlsym smoke coverage.

Validation

• Targeted SM100 pytest for GemmAmax AOT export/load and C++ smoke: passed.
• Targeted SM100 grouped GEMM dense/discrete AOT export/load regression: passed, 8 passed in 38.24s.
• DeepSeek DSv3 cudnn_jit_repro in Dockerfile-built container:
  • CUDNN_FE_AOT_MODE=write: passed, exported 6 .json, 6 .o, and 6 .so artifacts.
  • CUDNN_FE_AOT_MODE=read: passed, 0 cute.compile lines and completed from AOT artifacts.
  • CUDNN_FE_AOT_MODE=readwrite: passed, first process exported artifacts; second process emitted 0 cute.compile lines and loaded artifacts.

Notes

• The DeepSeek validation used the Dockerfile-built DSv3 container with local cudnn-frontend installed from this branch because the prebuilt NGC image required SSO access that was unavailable during setup.

[Anerudhan]

Lets wait till 1.25 branch cut happens before merging this in.

Thanks