[CuTeDSL] Decode UE5M3 top binade as finite Float32

[XFDG]

Summary

• decode scalar FloatNV8E5M3FNU to Float32 from its UE5M3 bit representation instead of using IEEE E5M2 widening semantics
• preserve finite values for the complete top exponent binade (codes 248-254)
• preserve 0xff as the format's sole NaN encoding
• add a 256-code CuTe DSL GPU regression test

Testing

• NVIDIA B200 (SM100), CUDA 13.1, nvidia-cutlass-dsl 4.8.0.dev0
• python -m pytest test_float_nv8e5m3fnu_conversion.py -q -rs (1 passed)
• the test verifies codes 0-254 exactly against the UE5M3 definition and verifies code 255 is NaN
• on the unpatched package with CUDA 13.1, the generic widening path fails NVVM compilation with unsupported operation; the patched bit-level lowering compiles and executes successfully
• ruff format --check python/CuTeDSL/_mlir_helpers/arith.py test/python/CuTeDSL/test_float_nv8e5m3fnu_conversion.py
• ruff check --ignore F403,F405 python/CuTeDSL/_mlir_helpers/arith.py test/python/CuTeDSL/test_float_nv8e5m3fnu_conversion.py (the ignored diagnostics are pre-existing wildcard-import diagnostics in arith.py)
• git diff --check

Contribution disclosure

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

Fixes #3602