Instead of mixing the CUDA driver API and the runtime API in the generated code,
use only the CUDA driver API.
GPU programs for CUDA gpuarray operations (except conv operations) are now
generated as a string that is passed to the python interface of libgpuarray.
libgpuarray then generates a cubin bytearray, which is embedded in the
generated code.  The generated code then uses the CUDA driver
API via the C++ interface of libgpuarray to load and launch the GPU program.

This has at least two benefits:

(1) This approach does not use the nvcc offline compiler to compile the
    generated code into the shared library.  It uses the host compiler
    directly, which is likely to be faster.  Note that, for cubin generation,
    libgpuarray still uses the nvcc offline compiler, but an improvement is
    being made to use NVRTC and ptxas instead of nvcc, which should be, again,
    faster.
(2) Mixing the CUDA driver API and the runtime API is typically discouraged.

Gpu reshape opt.

Fix test and better error message

[CRASH] fix crash on Raspberry Pi 1

Small visual changes (blue color) (bis)

Warn about pending ignore_border default value change.

Deactivate merge of assert as it cause cycle in the graph

Prod dimshuffle opt

Implement batched_tensordot in terms of batched_dot

Delete old stuff

Nouiz mixed