Use an ordered set for collecting operators to ensure deterministic
variable naming, as the variable name depends on the collection order.
(As an alternative I tried sorting the output of the collector by
string, but the stringification is very expensive on simple examples
like wave-min.)

Also, improve debuggability of any remaining bugs in loopy kernels
(such as any latent nondeterminism) by giving each assignment kernel a
separate name.

exec_mapper_factory argument to bind()

This change adds the ExternalCall node and execution support for it.
External functions may be passed as part of the execution context.

Avoid using specific Py3 sub-versions for CI

See merge request !34

Fix ToLoopyExpressionMapper for subscripted variables

See merge request !32

Caches for speedy bind

See merge request !30

Placate flake8 3.6

See merge request !28

Add Pytest/JUnit/Gitlab integration

See merge request !27

Preinstall pybind11 for pyopencl

See merge request !25

Use mpi4py.run in spawned MPI subprocessess

See merge request !24