one context per device.

Only scan is left.

This unfortunately renders pycuda unusable without a significant
amount of work which I am not ready to do right now. Since we have C
code for this op I just removed the python code.

- Convert GpuKernelBase to use contexts.
- Start using explicit context for all the ops.

This may break some stuff outside of basic_ops.py since sometimes a
context will be required, but should be ok otherwise.

Still in transitory state, this should work with ops that rely on type
context or default context (as long as they are equal).  Will
gradually move to type context only.