avoiding name collision and other things

also make sure it has the right sign

initial state can be one dimension less if the taps equal just [-1]

Make platform-dependent compiledir configurable

Cudandarray setitem

Minor stuff

This call cudaMemset under the hood, needed for Scan on gpu.

Although it probably works under Windows with /, using os.path.join is a better
way to define paths.

@pascanur fixed bug that made scan fail to compile on gpu in some cases

Bug subtensor merge

New scan [work in progress]