Its much faster to read the refs from 114 projects when the reader
is pure Python and just doing file IO than forking 114 git commands
and parsing their output.
The reader caches refs based upon file mtimes. If any single ref
file has been modified since the last read, we re-read the entire
repository's ref namespace. This simplifies the code as we don't
need to worry about shooting down symbolic-refs, but it may cause
more IO than is necessary if only one ref gets updated.
This change drops `repo branches` in Android from 1.658s to 0.206s.
Likewise, `repo sync` improves dramatically as well.
Signed-off-by: Shawn O. Pearce <sop@google.com>
