WIP: GPU tensors

[kohr-h]

Second WIP PR. Much earlier stage than #738 but maybe some comments already applicable.

TODOs:

• Make the basic structures (space, element, weighting) work
• Implement ufuncs and reductions
• Write optimized kernels for inner, dist and norm
• Fence imports and dependencies on pygpu
• Adapt the unit tests (complex dtypes etc.) and make them work for GPU arrays
• Do some benchmarking to find out thresholds for Numpy vs GPU
• Check if reductions in all axes run faster on flattened arrays. They should. No, because reshape of GpuArray objects copy the array, in contrast to Numpy.
• Update documentation
• Large-scale tests
• What to do with the doctests? They fail if pygpu is not installed. Should the package be mandatory?
  Update: manual doctests are fixed by adding a conditional in the __main__ section, need to fix runs by pytest also.
• Expose function to precompile kernels for linear space methods? Also add more of them.
• Keep gpuary_tensors.py namespace clean by making context initialization code a function only returning the relevant stuff.

[kohr-h]

From here on, the rest is currently copy-paste from npy_ntuples.py. Ignore it.

[adler-j]

Great news:

Theano/Theano#4705 (comment)

Just a note, we are doing conda packages to help install pygpu. This will
make this much easier to install on windows.

[kohr-h]

I know - but good to hear that they are trying to make it work for Windows, too.