Turns out that git v1.7 does not support `--date=uni`, but `%ct` should
be roughly equivalent.

Hexadecimals were used to save few characters, but it turns out that
while lexical sort should be fine, in version modes they do tend to
sort in strange ways (eg. `t5af5e80a < t5af5e800` in rpm).

Helps ensure that a later devel build will install over older one.

When decoupled from context of exit status, 0 and 1 can be confusing.

Essential fact.sh functions git_lasthash() and semver() have been updated
so that they explicitly depend on git or fail, ie. forget about the cache.

Previously git_lasthash() would try to use cache or fail, while semver()
would try to use cache or rely on mkit.ini, which in turn had to be hacked
by dist() in order to make git-less targets expand version correctly
(with build data, eg. `0.1.2+ge1f4a77.dirty`).

The old code was guilty of four sins: inconsistency (git_lasthash()
vs. semver()), side effects (semver() would actually write cache),
hidden dependencies (both actually relied on functions from mkit.sh)
and unnecessary complexity (speculative $version use in semver()).

New code remediates all of these by making both functions just do their
job (ie. get data from VCS).  Places that did benefit from the silent
caching are served by __cached() function created for this purpose.

To minimize code needed to import a module

Drop the unnecessary and space-wasting requirement that `do` and `then`
must be alone on line.  It's a thing of the past!

Decouple variable declarations so that each can have docstring.
In case of _relck() we also rename variable $relsrc to a more
context-specific term $oracle.

Where not required, they only add visual cruft.

although I haven't been able to put my finger on it, git-log must have
changed between Debian's 2.1.4 and Fedora's 2.5.something.  The previous
version worked on Fedora but not on Debian; hopefully this solution is
more robust.

Indeed a weird one...

Note that similar to __MKIT_PROJ_VERSION__, this value cannot be
reliably determined when outside git (e.g. distributor's build dir),
therefore must be kept inside dist source package.

Unlike version, though, user (ie. app developer, ie. mkit.ini
maintainer) most probably does not want to edit it, therefore we're
normally keeping it out of sight.

The value itself is normally just the last git commit's hash (SHA1 these
days), suffixed with ".dirty" mark if there were uncommitted changes
during build.  (Just the same as we do with semver())

If we're outside git and the hash could not be determined, that either
means the source package was mangled with or was built with older
version of mkit.  In either case, the value provided will be 'UNKNOWN'.

Some parts got a bit confusing over time, also there was repetition.

In this commit:

 *  the release fact gathering routines have been moved to separate
    module,

 *  some less useful routines have been removed,

 *  some internal interfaces have been changed to allow for more
    readable code.