Remove comm hash and add per-comm caches

[JDBetteridge]

It's broken and it's dangerous!

Now that we actually free our duplicated comms the hash_comm function is less than useless as the same id can be returned by different comms.

This PR also adds per-comm caches, which circumvents the need to hash a communicator in the first place.

Update:

This (necessarily) turned into a much bigger refactor than was intended. The key changes, beyond removing comm hash, are:

• Removal of Cached a base class used to create an object that would be cached. However, overriding the __new__ method effectively turned instances into singletons, but the objects didn't necessarily fit this design pattern. Moving forward only functions should be wrapped in a cache.
• Removal of @cached_property decorator as this is now part of the Python standard library.
• Refactor of load and get_so in compilation.py to fit with the new caching framework. This now compiles shared objects to RAM backed storage (/tmp) and only caches the .so to non-volatile storage ($PYOP2_CACHE_DIR).
• Add a PYOP2_SPMD_STRICT environment variable, for checking SPMD behaviour by enabling MPI.Comm.Barrier() on logically collective calls. (Switched off by default)

Moving forward (in pyop3???) we should also remove the ObjectCached base class as it suffers from the same issues as Cached, but this is another big refactor.

[connorjward]

procrastinating

[connorjward]

Looks fine.