Parsing `*aa bb* ` times N has strange performance characteristics about commonmark-hs

[jgm]

Profiling shows most of the time is taken by

               contents = take
                 (length (afters left) - (length (afters right) + 1))
                 (afters left)

in processEm.

[jgm]

After using a more efficient algorithm, we get:

% time python -c 'print("*aa bb* "*10000)' | commonmark >/dev/null
user	0m0.892s
% time python -c 'print("*aa bb* "*12000)' | commonmark >/dev/null
user	0m1.062s
% time python -c 'print("*aa bb* "*20000)' | commonmark >/dev/null
user	0m1.826s

Much better!