The first issue is unbound variable access for _updated_mirrors. This
requires new function that checks correctly that feed name is a key in
the associative array.
The second issue in the same function is that _updated_mirrors was never
actually set after feed was updated.

Another issue discovered and introduced during review is that
_git_mirror_lock can't call functions because in case of mirror it
spawns subshell where functions are not available. This simply reverts
all _git usages after _git_mirror_lock to git.

The another issue discovered was that repositories cloned using
git_checkout function were not actually using mirror. At the same time
the mirror URL construction is now placed in single function that can be
called from two locations it is used in.

There was also an issue with feed_url_replace function that was
invalidly handling URLs that had no reference. The effect was that the
URL for feed without any reference was handled by the first case and
result was something like https://example.org;https://example.org for
https://example.org

Another fix for mirrors was invalid detection of existence of the
mirror. The wrong variable was used and also the initial clone was
placed in the invalid directory.
There was also an issue if git mirror directory itself was missing. This
is fixed by simply creating it unconditionally.

And the last issue was with OpenWrt expecting full git history. This
actually does not break build but produces nasty warnings from git
commands that are run in script getver.sh that is called from top level
makefile. The fix here is to mask CLONE_DEEP when getting OpenWrt
repository.