[CuTe DSL] Cache each distinct one-time warning

[XFDG]

Summary

• make print_warning_once retain every distinct message instead of only the most recent one
• prevent an earlier warning from being emitted again after a different message evicts it
• add a regression test for the first, second, first sequence and isolate the global cache between tests

Testing

• B200 node, Python 3.12: an AST-extracted execution of the current-main method emitted 3 warnings for first, second, first and exited 2
• B200 node, Python 3.12: the fixed source emitted exactly 2 distinct warnings and exited 0
• python3 -m py_compile python/CuTeDSL/cutlass/base_dsl/dsl.py test/python/CuTeDSL/test_warning.py
• the source checkout was not imported as a complete DSL package because its generated cutlass._mlir extension is not built on the validation node

Contribution disclosure

AI assistance was used for repository research, implementation, and running the checks above.

Fixes #3501