CuTeDSL: print_warning_once re-warns when more than one distinct message is used (lru_cache maxsize=1)

Description

BaseDSL.print_warning_once (cutlass/base_dsl/dsl.py ~line 813) caches on (self, message) with @lru_cache(maxsize=1):

@lru_cache(maxsize=1)
def print_warning_once(self, message: str) -> None:

With more than one distinct message, every new message evicts the previous cache entry, so alternating messages re-warn forever instead of once each.

Repro against nvidia-cutlass-dsl==4.7.0:

import warnings
from cutlass.base_dsl.dsl import BaseDSL

dsl = object.__new__(BaseDSL)
with warnings.catch_warnings(record=True) as rec:
    warnings.simplefilter("always")
    for msg in ["A", "B", "A", "B", "A"]:
        dsl.print_warning_once(msg)
print(len(rec))   # 5 warnings emitted; correct behavior is 2 (one per unique message)

Actual output: 5.

Suggested fix

Keep an instance-level set of already-emitted messages, or move the cache to a module-level function keyed on the message alone with unbounded size.

Start in cutlass/base_dsl/dsl.py at BaseDSL.print_warning_once and reproduce the alternating-message example from the issue. Verify that repeated messages emit only one warning each, including when more than one distinct message is used; the existing reproduction provides the completion check.