v0.8: Mat-Muls on Nvidia Hopper and Blackwell

This release answers a few questions:

• CUTLASS vs CUBLAS performance: which to choose?
• How did MMA instructions change with Hopper H100?
• How did they change again with Blackwell B200?

Minor

• Add: Warp-Group Binary MMA (d6daf3a)
• Add: Larger m64n256k8 WGMMA variant (3e3530e)
• Add: Warp-Group Async kernels (6cc7e34)
• Add: f64 MMA PTX variant (ae450e5)
• Add: CuTe draft (fdea727)
• Add: CUTLASS placeholders (b1ab93d)
• Add: Hopper sm90a PTX kernels (4bcf74a)

Patch

• Improve: CUresult error handling (d74d430)
• Improve: Logging CUDA errors (953a696)
• Fix: Synchronize TCs (494ba52)
• Improve: Impossible %tid condition against NVCC (8a9c9c5)
• Make: Temporarily block CUTLASS (df1b39c)
• Improve: Cleaner PTX code (71dea0c)
• Improve: Avoid NVCC-specific features (3d65c7f)
• Fix: Re-creating a CUDA stream (e831650)
• Make: Compile in parallel by default (8e671c6)
• Make: Separate host-only code (f751fbf)
• Docs: Counter-intuitive PTX facts (822fa2f)
• Docs: H200 vs MI 300X vs GB200 specs (cc36bcd)
• Make: CUTLASS dependency (f272c40)
• Fix: Synchronize cuBLAS for profiling (4077f26)
• Docs: Blackwell tensor cores (ec35b35)
• Fix: Missing _Float16 in NVCC, use half (71cadca)
• Improve: Same size range for GEMM (d914fce)
• Fix: Different output size for cublasGemmEx (304c880)