Optimise line splitting in shell

[Chris Warburton]

We split long lines, but the way we're doing it is wasteful according to
Emacs's profiler. Start profiling with:

  M-x profiler-start

Do some stuff in the terminal, then do:

  M-x profiler-report

It looks like our shells are mostly slow for 2 reasons: ANSI colour
coding and line splitting. The ANSI colour coding was pretty bad, using
Emacs's built-in functionality, so I've switched to xterm-color instead
and it seems to be faster. Now there's about an equal amount of CPU
spent on colouring and splitting. There's no need for splitting to be so
slow.

In particular, we might want to figure out what functions are written in
C and what's Elisp. If we can e.g. check whether a string contains a
newline using a C function, then we should probably use that on chunks.
Yet if that function is Elisp, we might be better off just looping
through the characters ourselves.

Also watch out for Schlemiel the Painter algorithms, e.g. appending over
and over when we should be using a 'builder'.