fix: disable tensor cores for FP32 backward GEMMs — NaN corruption (ALB-076)

[noahgift]

Problem

CUBLAS_GEMM_DEFAULT_TENSOR_OP (algorithm 99) produces ALL NaN output for transposed backward GEMMs (Trans/NoTrans, NoTrans/Trans) when input gradient magnitudes reach ~1e5. This occurs around block 18 of a 24-layer backward pass where gradient magnification grows from ~1e-5 to ~1e5.

Forward GEMMs (NoTrans/NoTrans) are unaffected.

Five Whys

1. Why NaN weights? → optimizer reads NaN gradients
2. Why NaN gradients? → cuBLAS backward_a/b output ALL NaN
3. Why NaN output from valid inputs? → tensor core GEMM algorithm
4. Why only backward? → backward uses Trans flag, forward doesn't
5. Why only after ~5 blocks? → gradient magnification reaches ~1e5

Fix

Switch from tensor core math to SIMD math:

• CUBLAS_TF32_TENSOR_OP_MATH → CUBLAS_DEFAULT_MATH
• CUBLAS_COMPUTE_32F_FAST_TF32 → CUBLAS_COMPUTE_32F
• CUBLAS_GEMM_DEFAULT_TENSOR_OP → CUBLAS_GEMM_DEFAULT

Performance

cuBLAS SIMD is still 6-14x faster than hand-written PTX:

• PTX baseline: 890 tok/s, 2.6% MFU
• cuBLAS SIMD: 5,216 tok/s, 15.1% MFU