This patch adds caching for both layouts and for HarfBuzz face objects.
The granularity of the cache for layouts is words, so it splits the
input string at word boundaries (using a heuristic). There are is also
some refactoring to reduce the amount of allocation and copying, and
movement towards properly supporting contexts.

The size of the caches is a fixed number of entries; thus, it is
possible to consume a large amount of memory by filling the cache with
lots of large strings. This should be refined towards a scheme that
bounds the total memory used by the cache.

This patch fixes bug 15237293 "Regression: Measure performance is
significantly slower with minikin".

Change-Id: Ie8176857e2d78656ce5479a7c04969819ef2718d