diff: avoid useless filespec population

diff: avoid useless filespec population

builtin_diff calls fill_mmfile fairly early, which in turn
calls diff_populate_filespec, which actually retrieves the
file's blob contents into a buffer. Long ago, this was
sensible as we would need to look at the blobs eventually.

These days, however, we may not ever want those blobs if we
end up using a textconv cache, and for large binary files
(exactly the sort for which you might have a textconv
cache), just retrieving the objects can be costly.

This patch just pushes the fill_mmfile call a bit later, so
we can avoid populating the filespec in some cases.  There
is one thing to note that looks like a bug but isn't. We
push the fill_mmfile down into the first branch of a
conditional. It seems like we would need it on the other
branch, too, but we don't; fill_textconv does it for us (in
fact, before this, we were just writing over the results of
the fill_mmfile on that branch).

Here's a timing sample on a commit with 45 changed jpgs and
avis. The result is fully textconv cached, but we still
wasted a lot of time just pulling the blobs from storage.
The total size of the blobs (source and dest) is about
180M.

  [before]
  $ time git show >/dev/null
  real    0m0.352s
  user    0m0.148s
  sys     0m0.200s

  [after]
  $ time git show >/dev/null
  real    0m0.009s
  user    0m0.004s
  sys     0m0.004s

And that's on a warm cache. On a cold cache, the "after"
case is not much worse, but the "before" case has to do an
extra 180M of I/O.

Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>