Also, use output memory even if not c-contiguous, instead of crashing.

The C code assumes that the output memory buffer is C-contiguous,
and it would be work to change, so we re-allocate a C-contiguous out.

Also make sure the output memory layout is suitable, otherwise it would
be copied a second time.

Gpu stuff

This affect all the function that call this inner function.

Check that Alloc nodes are not inserted during opt

Gpu setsubtensor

This is needed for GpuAdvancedIncSubetensor1

making the subtensor merge optimization work

Fix for CUDA 4.1 on OS X, prevent -no_pie flag

Create the $TMPDIR directory if needed.

Always introduce Alloc node when calling alloc()

Otherwise, nvopencc complains.

Also use correct mode in test.