This is a significant speed up with the gc as most of the time, we
don't need it and allocating on the GPU is slow.

C code gpusub

Also add missing decref in error handling code.

Revert "Optize GpuSubtensor by 11% by moving the reversed call in the make node instead of the perform."

This reverts commit aeedaa73.

We will reuse the cpu Subtensor c_code, so we need the parameter to be
in the same order.

So we use PyArray_BYTES instead of PyArray_DATA

-use range instead of a tensor with the same values
-small refactoring to make the code easier to read.

I'll reuse it for gpusubtensor.

Small fix

Np17

NEWS:

documented, speed up(50% and renamed _dfs_toposort (now _contains_cycles)
Rename FunctionGraph.nodes -> apply_nodes. The old nodes still work, but get warned if used.
Made a parent class called Node for Apply and Variable class that are both FunctionGraph nodes

Use the async gpu kernel call by default.

make is clear that the allocated space is not needed.