Reimplement xcounter based on pre-rendered pixmpas.

This alters the xcounter idea, instead of caching XmStrings and doing
full-blown font rendering with each update, we can pre-render all the
needed strings to a pixmap and stitch them together using XCopyArea.

This achieves basically the same visual result but clears are eliminated
completely which should further reduce the chances of visible flicker.

While we are rejigging this, tune the state structure to be much
more focused to the actual usage, preferring fixed limits and compact
data representation over dynamic allocation.