Reland [Performance] Speed up ConvertUtf8ToUtf16Internal

Reland change: Only fast-path on aligned words. Performance actually
seems a little improved with this change.

Original commit message:

This is a hot function that is made much faster by checking for
sequences of ascii-characters a word at a time for fast copying. It
falls back to the original 1-character at a time code.

This change reduces CPU usage by .54% on a busy page, as described in
the bug.

The primary cleanup in the 1-character codepath is that rather than
keeping track of an original span and cursor and revert the cursor where
necessary, we just move the span as we use it.

Bug: 44107248
Change-Id: I3095a0bd1b4fab50963e6f6783f7b81d338bb5a5
Reviewed-on: https://chromium-review.googlesource.com/c/chromium/src/+/6915475
Reviewed-by: Nico Weber <thakis@chromium.org>
Commit-Queue: Josh Karlin <jkarlin@chromium.org>
Cr-Commit-Position: refs/heads/main@{#1511059}