Remove obsolete SM100 fully fused KDA prototype

[icavan]

Summary

• remove the obsolete kda_fully_fused_wip.py SM100 prototype and its dead-path wrapper
• remove the corresponding fused-forward test and benchmark
• remove the unused get_kda_fused_fwd dispatcher
• clean up README, repository-layout, benchmark commentary, and Ruff references

Why

The experimental SM100 fully fused path is not part of the active KDA implementation, but it retains roughly 6,000 lines of WIP kernel code plus wrapper, test, benchmark, dispatch, and lint-maintenance surface. Removing the complete dead path avoids stale imports and makes the supported modular/SM90 prefill paths clearer.

The maintained SM90 prefill coverage remains in tests/test_kda_sm90_prefill_vs_fla.py and tests/test_kda_sm90_intracard_cp.py.

Impact

This removes the unpublished experimental SM100 fused entry point. Supported modular SM100 KDA and maintained SM90 prefill implementations are unchanged.

Validation

• pre-commit run --all-files
• python -m compileall -q cula/utils.py benchmarks/bench_kda_sm90_prefill.py
• repository-wide scan confirms no remaining references to the removed module, wrapper, test, benchmark, or dispatcher
• git diff --check

GPU tests were not run because the configured B200 host aistudio-58650011-ssctl could not be resolved from the current environment.

[icavan]

Post-#119 validation

Validated PR #116 against the latest main after #119 merged.

• GitHub's synthesized merge commit is ea452466fbf9d46c84355bf4727d6e04a98f4c3c, with parents 5370b14637c3bb4bc4242b334e6b7b00a05c6e41 (main including Detect tcgen05 load/store API by signature #119) and b1bc3aefebbb25d67a531f5c1008302ec106a921 (PR Remove obsolete SM100 fully fused KDA prototype #116).
• GitHub reports the PR as mergeable; the current blocked state is the required-review gate, not a merge conflict.
• There are no overlapping changed paths between Remove obsolete SM100 fully fused KDA prototype #116 and the commits added to main since Remove obsolete SM100 fully fused KDA prototype #116's original base.
• The synthesized merge contains no remaining code references to the removed WIP module, wrapper, benchmark, test, dispatcher, or sm100_fused package.
• pre-commit run --all-files, git diff --check, and AST parsing of all 120 Python files passed on the synthesized merge.
• On the remote B200/SM100 environment with CuTeDSL 4.5.3 and FLA v0.5.0, the broader SM100 KDA run completed with 210 passed and 174 skipped. The 18 failures were all environment-only imports of the absent compiled module cula._cudac_sm100 in the isolated checkout; no test failed inside the Remove obsolete SM100 fully fused KDA prototype #116 deletion or Detect tcgen05 load/store API by signature #119 compatibility paths.
• An isolated extension rebuild was attempted to remove that environment limitation, but the remote CUDA development image lacks cusparse.h. This is unrelated to either PR and prevents completing the extension-backed kda_fast cases in that isolated checkout.

Conclusion: #119 introduces no conflict or follow-up requirement for #116. No source update or rebase-only commit is needed; PR #116 is ready for normal review/merge from a code-integration perspective.

[cherhh]

LGTM

[tongke6]

LGTM