Avoids overhead due to readdir calls while searching a pack index. ok op@

with lots of help from stsp for the manpage bits!

ok op@

Simplify some phrasing, explain what preconditions `got patch' has and what happens to the work tree when an error occurs.

Don't allow `got patch' to delete files that are not known, or add files that are already known and to edit files that are known, not obstructed and without conflicts.

These commands perform potentially long-running operations on the repository after reading information from the work tree. There is no need for them to keep the work tree locked until the end. Doing so blocks other commands the user may want to run concurrently. For example, the user may want to be able to run 'got diff' in the work tree while browsing 'got log' output in less(1). Pointed out by Misha on gameoftrees IRC.

add two helper functions (schedule_add, schedule_del) and move the guts of apply_patch into a new function `patch_file'. This simplifies apply_patch and makes easier to figure out what happens. Then, drop GOT_ERR_PATCH_PATHS_DIFFER since we allow the to rename files.

otherwise got-read-patch can't parse any patch because the position indicator is at the end of file. While here fix the return too and set patchfd to -1 on error (stsp suggestion.) ok thomas, stsp

ok thomas, stsp