Scan grad speedup bug

support for multiple axes in MaxAndArgMaxOp

Fix crash introduced by gh-3476

Scan backend speedup

Don't be more strict then numpy rtol on float64.

Multiple fixes preparing for multi-gpu

Fix arange to avoid overflow.

Better error message

Rework amdlibm exclusion on GPU.

Otherwise it is possible to get into a situation where we use the dnn flags, but the cuda ones aren't defined and stuff breaks.

Also starts adressing issue #3468 a bit.