ProcessLogger: avoid spawning threads in noninteractive environments (https://github.com/firedrakeproject/firedrake/issues/1422)

Fix a misleading assertion in test_pickling_with_module_import(). Only globals in the module are pickled

module imports.

For portability to old Python versions, add a backported version
of importlib.

Closes #1

* Add a CacheCollision warning subclass.
* Track stack level so that warnings show user code line numbers.

Example:

Modifying line 287 of test_persistent_dict.py so that the warning
isn't captured by the test, we get this:

```
test_persistent_dict.py:287: CollisionWarning: pytools-test: key collision in cache at '/tmp/tmpq9p3i0b9' -- these are sufficiently unlikely that they're often indicative of a broken implementation of equality comparison
  pdict[key2]  # user code
```

The old behavior was:

```
/home/matt/src/pytools/pytools/persistent_dict.py:466: UserWarning: pytools-test: key collision in cache at '/tmp/tmpuk8js9jw' -- these are sufficiently unlikely that they're often indicative of a broken implementation of equality comparison
  % (self.identifier, self.container_dir))
```

This adds an optional in-memory LRU cache. To use the cache, you
supply a *in_mem_cache_size* parameter to the PersistentDict.

In order to properly support cache invalidation this also implements
version tracking of files. This change is backwards compatible with
existing persistent dict caches.

test_spatial_btree currently fails if numpy is not available.

test_p_convergence_verifier uses the pytest importorskip function
to skip the test when numpy is not available.

Do the same for test_spatial_btree.

disable kwargs support of memoize by default, remove iteritem wrapper since it will be handled by 2to3

add keyword argument support and key= keyword argument to function memoize in order to support arguments that are not directly hashable
add test for pytools.memoize