Zero-initialize all fields of Memcpy*D

Some fields were left uninitialized, which doesn't seem very Pythonic.
The CUDA_MEMCPY3D structure is now value-initialized, which ensures all
its members are zero.

Add ability to get handle to CUDA Arrays

Added function to create CUDA Arrays directly from GPUArrays

Just to complete, added function to create CUDA Arrays directly from GPUArrays, maybe in future Texturea Objects and Surfeces Objects could be supported ..

extFPTextureSurfaces

Implemented function to make Cuda Arrays from numpy 2D or 3D array including complex and double support

Improve error handling of non-array types in ArgumentHandler

Ensure the test actually corresponds to the cuda code.

Add support for passing shapes that are ndarrays to GPUArray constructor

Add ability to set real and imaginary components of complex numbers

Add a GPUArray.squeeze() method.

Sometimes useful in cases where numpy and GPUArray shapes
don't match due to length 1 dimensions. This can occur
when slicing GPU and CPU arrays and helps to suppress warnings
about non-matching array shapes.

Fix stream parameters in asynchronous methods.

Within gpuarray.set_async, the stream keyword
argument in the call to gpuarray.set was
always set to None.

  return self.set(ary, async=True, stream=None)

This keyword argument is now always set to value
supplied in the arguments to gpuarray.set_async.

  return self.set(ary, async=True, stream=stream)

Within _memcpy_discontig, Memcpy2D.__call__
was being called for asynchronous cases as follows:

  copy(stream=stream)

However, this does not match the C++ function
definition. Instead, they are now called as follows:

  copy(stream)