The pattern /(.{1,N})(\s+|$)/ runs in quadratic time on long
no-whitespace lines, tripping Ruby 3.2's Regexp.timeout. In
production this surfaces when HTML emails contain long URLs,
encoded tracking links, or wide table cells flattened to a
single line.

Replace the gsub with a linear loop that scans a window of
line_length+1 chars for the rightmost whitespace and falls
back to a hard break when none exists. Behavior matches the
original on normal inputs (the existing 'wraps lines' spec
still passes) and stays linear on pathological inputs.

Fixes the production crashes tracked at
https://kit-qq.sentry.io/issues/KIT-RAILS-WV

The previous version split any line longer than line_length on a
strict char boundary when no whitespace was found within the window.
That regressed the gsub original behavior, which left long no-whitespace
runs alone (the regex simply didn't match), so URLs and other long
tokens stayed on a single line and remained clickable in plain text
clients that rely on whole-line URL auto-detection.

Now the else branch scans forward to the next whitespace and breaks
there, falling back to leaving the line intact if no whitespace exists
at all. Linear time is preserved.

Spec coverage updated: pathological no-whitespace input now asserts
the token is returned intact, and a new case covers a long URL
surrounded by short words.

The old gsub ended with .strip, dropping any leading or trailing
whitespace from the wrapped result. The new implementation was
omitting it, which only showed up under direct callers of
custom_word_wrap (convert_to_text scrubs whitespace later, so
end-to-end behavior was unchanged). Adding it back keeps the
public contract identical to pre-refactor.

ECO-5370: Fix Regexp::TimeoutError in custom_word_wrap