Add tileiras optional dependency for bundled compiler support

[hannahli-nv]

Summary

Addresses #79: users without a system CUDA Toolkit installation need cuda-tile[tileiras] to compile TileGym kernels at runtime.

• Add tileiras extras_require to setup.py, enabling pip install .[tileiras]
• Restructure README installation section with Option A (bundled tileiras via pip) and Option B (system tileiras)
• Add comment in requirements.txt about the tileiras option

Closes #79

Details

TileGym kernels are compiled at runtime by the tileiras compiler from cuda-tile. Users can either:

1. Install the compiler via pip: pip install .[tileiras] (bundles nvidia-cuda-tileiras, nvidia-cuda-nvcc, nvidia-nvvm into the Python env)
2. Use a system CUDA Toolkit 13.1+ which includes tileiras

Previously, the README only documented pip install . which installs cuda-tile without the compiler binaries, causing failures for users without a system CUDA Toolkit.

Test plan

• Verify pip install .[tileiras] installs nvidia-cuda-tileiras and bundled compiler
• Verify pip install . still works on systems with CUDA Toolkit 13.1+
• Verify benchmarks run successfully with bundled tileiras

CI Configuration

config:
  build: true
  test: ["ops", "benchmark"]

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.