Add parallel-safe LRU cache and global cache manager objects

[connorjward]

There are occasions in PyOP2/Firedrake where we repeatedly perform expensive operations but are unable to cache them because:

• No suitable object exists upon which the object may be cached, and
• The object references "large" data structures

A classic example of this is something like computing the adjoint of a form. Forms are often non-persistent (e.g. solve(u*v*dx == f*v*dx, ...)) and they hold references to the coefficients so a global cache is not suitable. Computing adjoint is a significant cost when computing the derivative of adjoint problems so caching results is desirable.

This PR introduces parallel-safe LRU caches. I believe that this would address the problems described above.

I have also introduced the concept of a "cache manager" (#693) to facilitate cache cleaning and inspection.

Closes #696 #693

TODOs:

• Code cleanup and documentation
• Write tests
• Enable configuration of cache sizes via environment variables - users with memory constraints would want smaller caches

[connorjward]

@JDBetteridge: we should have pretty printing of cache statistics.

[connorjward]

I need to tie these caches to the communicators so they are cleaned up at the right time (in case we spew lots of communicators and leak that way).

[connorjward]

Closing as @JDBetteridge has already merged a much nicer version of what I was trying to accomplish.