Version 15: Lower tl.dot to CUDA cores

[pbelevich]

Summary

• lower staged tl.dot operands to a per-thread CUDA-core FMA loop
• synchronize shared-memory reads before runtime K-loop tile reuse
• execute correct single-tile and multi-K-tile matrix multiplication on CUDA
• document the working tiled matmul path and Version 16 boundary

Testing

• make (289 passed)