Fix cuDNN convolution precision on Ampere+ GPUs

[joelnn]

On Ampere and later GPUs (SM 8.0+), cuDNN's default math mode permits TF32 Tensor Core operations which use reduced mantissa precision. This causes numerical differences when comparing CUDA vs CPU convolution results, particularly in cudnnConvolutionBackwardFilter().

Explicitly set CUDNN_FMA_MATH to force true FP32 computation for consistent numerical results across all GPU architectures.

[davisking]

Sweet, thanks for another PR :D

[Cydral]

@joelnn, Would it be worth considering CUDNN_TENSOR_OP_MATH_ALLOW_CONVERSION instead of CUDNN_FMA_MATH to maintain Tensor Core performance on Ampere+ GPUs? This would only require relaxing one or two test tolerances from 1e-3 to 2e-3 in test_conv(), which seems acceptable given the significant performance benefit...

[joelnn]

@Cydral thats alright with me, I mainly wanted tests to pass. I would've thought that reducing precision should be opt-in, but following the default policy of cuDNN would also be a reasonable policy.