I've now written a similar function to revert the buffer, but using
replace-buffer-contents as I was suggested.  The new attached file
contains such function, which I called revert-buffer-with-fine-grain, to
honor a previous post.

I kept it as similar as I could to the revert-buffer-by-hunks
function, and went ahead to time both functions (I figure there will be
enough time to add functionality to them).  I additionally timed
revert-buffer.

For the files in Bug#31888, the results were:
revert-buffer: 0.122238819
revert-buffer-with-fine-grain: 2.85150876
rbbh-revert-buffer-by-hunks: 2.044583634

all times in seconds, substracting GC time.

To gather more data, I compared emacs-25.3 and emacs-26.2 C source
files (from the src/ directory), checking them out from the git
repository.  I did the test as if I was going back from emacs-26.2 to
emacs-25.3.  I think these files provided a good range of changes in
size and number of hunks, obtained from `diff --normal'.

I attach the data as ascii file.  Let me know if I should send it in
another format.  The tests are with the latest master, optimized build,
run as `emacs -Q'.

For the majority of files, revert-buffer and
revert-buffer-with-fine-grain spent a similar amount of time when the
buffers or changes are not too big, so I believe that's promising for
implementing this functionality.  On the contrary, reverting by hunks
(by calling diff), takes 1 to 4 times more than the other functions, in
these cases.

When there are many changes, replace-buffer-contents starts taking too
much time (as the docs warns).  In those situations, reverting by hunks
does a much better job than using replace-buffer-contents, but I'm not
sure if it does a good enough job.

Some problematic files:
doprnt.c:
revert-buffer-with-fine-grain took 28.6 times more than revert-buffer
and rbbh-revert-buffer-by-hunks took 1.4 times more.

alloc.c:
revert-buffer-with-fine-grain took 18 times more than revert-buffer
and rbbh-revert-buffer-by-hunks took 6.5 times more.

lisp.h:
lisp.h took almost 9 minutes to revert, using
revert-buffer-with-fine-grain.  It took 0.86 seconds to revert with
rbbh-revert-buffer-by-hunks.

I'll wait for opinions about the data I've collected.  At least it was
super fun writing these tests.

Best regards,
Mauro.