Add pyproject.toml for modern Python packaging #5

Problem

The project uses a legacy setup.py as the sole packaging configuration. Modern Python packaging uses pyproject.toml (PEP 517/518).

Also, setup.py hardcodes nvidia-cutlass-dsl==4.4.2 as the only dependency — no version range flexibility.

Recommendation

Add a pyproject.toml with:

• Build system declaration (setuptools + CUDA extension)
• Project metadata (name, version, description, license, Python version)
• Dependencies with version ranges (e.g., nvidia-cutlass-dsl>=4.4,<4.5)
• Optional dev dependencies (pytest, mypy, ruff)

Keep setup.py only for the CUDA extension build logic, or migrate fully to pyproject.toml + a build script.

Impact

Modernization. Enables pip install -e . to work cleanly, makes dependency resolution smarter, and follows current Python packaging standards.