Fix the last remaining problems with the existing clients and make the kernel code for MRG more OpenCL-friendly.

Each item will be precompiled separetly and embedded into the c_code
of the Op. This allows ops that need multiple kernels or that will
choose between alternatives at runtime to use this interface.  It also
groups all kernel-related parameters under one object.

This change also save the source of the kernel code to re-attempt
source compilation in case the binary is rejected for some reason
(some implementations do not support reloading from pre-compiled
kernel).

There may still be more changes to how stuff works under the hood
(most notably a blacklist of bad runtime/drivers that crash when
attempting to create a kernel from a binary), but the visible
interface should not change anymore, so now is the time to start using
it more.

Follow the rename of libgpuarray.

gh-1816 PR with small fixes

Remove import of device_properties at the top of the file, as this break import of Theano when there is no GPU available.

This is needed for the new gpu op tests to pass.

Corrected check for whether cuda is active or not and compile only one addition kernel instead of many in the perform() method.

Mostly adapted Op and tests to new backend. TODO: Remove faulty python implementation from _dev20 version of op

Copied GpuAdvancedIncSubtensor1() and GpuAdvancedIncSubtensor1_dev() ops and their tests to sandbox/gpuarray

Fast rng mrg

Conflicts:
	theano/sandbox/rng_mrg.py
	theano/sandbox/test_rng_mrg.py

Corrected check for whether cuda is active or not and compile only one addition kernel instead of many in the perform() method.

Cuda ndarray fix

Speed up from 0.085s to 0.055s in mrg init.