Earlier when we read a tree into a temporary index, we read it
from scratch. Start from the current index and use read-tree -m
to preserve cached stat information as much as possible, in
order to speed up "git add -u". This makes "git stash" usable
in a source tree of nontrivial size.