(This failed mainly because numpy's object arrays don't
participate in cyclic gc.)

This was moved from hedge, and while moving, I added data on devices
supporting Compute Capability 1.2 and above.

Conflicts:

	Makefile.in
	src/python/gpuarray.py

added some changes from the main repository to keep it more or less compatible. I also deleted all references to the simple array since it's not longer needed. The GPUArray initializes it self or you can specify the device number as last argument.

Refactor CUDA tracing.
Fix new-style error reporting.
Only release GIL for certain, non-cleanup functions.
  (Cleanup functions may be called after Python
  deinitializes, which then causes the GIL release code
  to abort.)
Add sticky allocators to GPUArrays.