Calling out to glom too often hurts performance, and before this patch
we did so to check for key membership AND access the value for a given
key, with no caching in between. This also adds that if a key doesn't
exist, we'd end up in exception handling, which is notably slow in
Python.

This patch applies an optimization for the simple cases in a couple of
ways:

* we store which (nested) keys exist. This happens by setting the key
  in an internal set whenever a write operation to that key happens. If
  the FormioData is initialized with a data structure, the UserDict
  internals cause the __setitem__ method to be called, which ensures
  the keys is added to the set.
* if the key doesn't contain the '.' character, we know that it must
  be a top-level/root key, so we can shortcut the membership test by
  checking for membership in 'self._keys' rather than hitting the
  glom code path
* in __getitem__, we can also treat simple keys as a simple case and
  directly access self.data, rather than going through glom for this

Note that this patch doesn't handle deleting keys - __delitem__ is not
implement, so the key removal doesn't either. I also suspect there
may be some weird cases with 'data.update(...)' variants where nested
keys are changed and stale data may be left in the _keys membership,
so that's why I'm erring on the side of caution and only consider
root-level keys for the shortcut code path.

This patch can definitely be improved upon, but it must also easily be
backported and a larger refactor can be done in a separate PR.

(cherry picked from commit 7b6432ea5377a4fe0f119a6b55f49893c62f62a6)