This branch is where I (jmarc) try to implement string_wise metrics
computation. This is done through a series of cleanups. The expected
speed improvement will only be visible at the end of the road: indeed
for now we intend to keep unchanged behavior for testing purposes.

What is done:
* Make TextMetrics methods operate on Row objects: breakRow and
  setRowHeight instead of rowBreakPoint and rowHeight.

* change breakRow operation to operate on text strings on which
  metrics are computed. The list of elements is stored in the row object
  in visual ordering, not logical.

* re-implement cursorX using row elements

* Implement proper string metrics computation (with cache), when
  lyxrc.force_paint_single_char is false.

Next steps:

* re-implement getColumnNearX using row elements

* get rid of old code of cursorX and getColumnNearX (which have been
  kept for comparison purpose).

* re-implement row painting using row elements (can it be done?)

* profile and see how performance can be improved.

* Document the code

Difference in behavior
* end of paragraph markers metrics are computed with the font of the
  actual text, not default font. This will be extended to the other
  methods.

* When cursor is after a LtR separator just before a RtL chunk, the
  cursor posiiton is computed better with the new code.

Other differences that should be considered as bugs
* words longer than the screen are no monger broken at an arbitrary
  point. This is a problem for languages like chinese that do not use
  separators.