Do simple in-index merge with diff+apply instead of read-tree

The advantage is that patch application will resolve some file content
conflicts for us, so that we'll fall back to merge-recursive less
often. This is a significant speedup, especially since merge-recursive
needs to touch the worktree, which means we have to check out the
index first.

(A simple test, pushing 250 patches in a 32k-file repository, with one
file-level merge necessary per push, went from 1.07 to 0.36 seconds
per patch with this patch applied.)

Signed-off-by: Karl Hasselström <kha@treskal.com>