Make the docstrings-wide check 70x faster

Make the docstrings-wide check 70x faster

Instead of performing a number of expensive transformations on the
original doc string and then use a dynamically-created regexp to find
wide lines, step through the lines in the unmodified string and only
perform the transformations on lines that exceed the limit.
This is sound because the transformations are contractive.

The new check will usually not cons nor perform any regexp matching.

* lisp/emacs-lisp/bytecomp.el (bytecomp--docstring-line-width): New.
(byte-compile--wide-docstring-p): Cheaper implementation.