cublas error with "--math-mode=fast"

[kbarros]

From the terminal:

$ julia --math-mode=fast

(v1.1) pkg> status
    Status `~/.julia/environments/v1.1/Project.toml`
  [c5f51814] CUDAdrv v4.0.4
  [be33ccc6] CUDAnative v2.5.5
  [3a865a2d] CuArrays v1.4.7
  [...]

using CuArrays
CuArrays.CUBLAS.handle()

gives me this error:

ERROR: MethodError: Cannot `convert` an object of type Ptr{Nothing} to an object of type CuArrays.CUBLAS.cublasMath_t
...

The simple fix is to reorder arguments in this line:
cublasSetMathMode(handle, CUBLAS_TENSOR_OP_MATH)

However, the existence of this error seems to indicate that nobody is relying on the behavior that "fast-math" mode enables CUBLAS_TENSOR_OP_MATH. I would suggest making the "fast-math" independent of CUBLAS_TENSOR_OP_MATH by deleting the highlighted code here:
https://github.com/JuliaGPU/CuArrays.jl/blob/master/src/blas/CUBLAS.jl#L55-L59

If the user wants CUBLAS_TENSOR_OP_MATH, then they really want it, and they can make that explicit by calling:
cublasSetMathMode(handle(), CUBLAS_TENSOR_OP_MATH)

[maleadt]

Sorry this took so long. Fixed in #567. I think leaving that code in makes sense, since it conforms to other changes when running under fast math mode. Downstream users can still set the mode explicitly if they want to (cc @MikeInnes, maybe Flux wants to).