fix: improve build error messages and add architecture status output

[yurekami]

Summary

This PR improves the build experience by providing clearer error messages and better feedback during compilation.

Build System Improvements

• Replace cryptic assert with descriptive RuntimeError messages
• Add actionable guidance when CUDA version is too old for SM100 (Blackwell)
• Show which GPU architectures (SM90/SM100) are enabled during compilation
• Add error handling for missing or broken CUDA installations
• Prevent silent failures when no architectures are enabled

Example: New Error Message for CUDA < 12.9

Before:

AssertionError: sm100 compilation for Flash MLA requires NVCC 12.9 or higher...

After:

RuntimeError: SM100 (Blackwell) compilation requires CUDA 12.9 or higher, but found CUDA 12.4.

Options to fix this:
  1. Upgrade to CUDA 12.9+ to enable SM100 support
  2. Disable SM100 compilation by setting: FLASH_MLA_DISABLE_SM100=1

Example: FLASH_MLA_DISABLE_SM100=1 pip install -v .

Note: SM90 (Hopper) kernels will still be compiled and available.

Python Code Style

• Use assert not causal instead of assert causal == False (PEP 8 E712)
• Improve assertion error message clarity

Test plan

• Verify error messages display correctly with CUDA < 12.9
• Verify compilation succeeds with FLASH_MLA_DISABLE_SM100=1
• Verify architecture status is printed during build

Relates to #79

🤖 Generated with Claude Code