Make PatternSub register the proper trackings and fix a bunch of call sites of local_optimizer to use the proper calling form.

Most of the tests pass now, except a bunch in the Cuda backend.

- Fix existing optimizers in base code to register properly.

Compile the inner fct in scan only once per instance.

Conv3d2dinplace

Add config option to specifiy cublas library name.

Assorted gpuarray work.

[MRG] Fix crash when specifing march in cxxflags.

GpuEye compyte

Use a ternary op since it seems the cuda compiler cannot resolve overloads by going through simple typedefs.

Cleanup

Refactor it to make a registry op c code.

The new Gpu back-end reuse those 2 fct.

Convert GpuEye to be a real compyte op.  This is, where possible, a direct translation of the original code and does not reflect the preferred style.

add assert too.

They are used by conv3d2d.