Detect tcgen05 load/store API by signature

[icavan]

Summary

• replace the CuTeDSL version gate with capability detection based on the generated tcgen05_ld and tcgen05_st function signatures
• detect the optional num argument independently for load and store, and select the supported store value keyword (r or val)
• fail fast with an explicit error when an unknown or ambiguous generated binding is encountered
• add unit coverage for the legacy, inferred-shape, mixed, and unsupported signatures
• migrate the SM100 intracard CP tests from the removed private _no_cp guard to the public use_intracard_cp=False control

Motivation

The low-level generated NVVM binding changed between CuTeDSL 4.5.2 and 4.5.3: num was removed and the store value keyword changed from r to val. A package-version comparison is brittle because this is an internal API and distributions may backport or ship the binding change independently. Inspecting the callable interface directly makes the compatibility decision match the API that is actually installed.

The intracard CP tests also retained two calls to the private _no_cp keyword after production dispatch was migrated to the shared CPMode interface. Using use_intracard_cp=False restores the intended serial baseline and fallback behavior without reintroducing a second, conflicting dispatch control.

Validation

Validated on an SM100 GPU in isolated environments:

• CuTeDSL 4.5.2 detected as ld_has_num=True, st_has_num=True, st_value_keyword="r"
• CuTeDSL 4.5.3 detected as ld_has_num=False, st_has_num=False, st_value_keyword="val"
• tests/test_cutedsl_compat.py: 4 passed
• tests/test_ptx_umma_ws.py::test_ws_ss_tf32: passed on both 4.5.2 and 4.5.3
• KDA forward/backward smoke test, including TMEM load and store: passed on both 4.5.2 and 4.5.3
• KDA fast group: 18 passed, 38 skipped, 136 deselected
• pre-commit run --all-files: passed
• Ruff, formatting, bytecode compilation, and git diff --check: passed

Related to #118.

[cherhh]

LGTM.

[zheyang0825]

lgtm