This should speed up CI a bit.

CI stages are separated with a barrier, i.e. all the jobs from the
previous stage must finish before first job of a next phase starts.

In our current build phase we have the following jobs:
 * build (gcc)
 * build (clang-sanitizers)
 * clang-tidy

The build jobs are kind-of fast (ca. 5 and 10 mins, but not in parallel,
i.e. 15 mins in total) and the clang-tidy job is slow (ca. 27 mins).
After the clang-tidy jobs finishes we move on to the test phase where
the CI has to download dependencies (i.e. the build directory) in order
to run tests (ca. 3 + 5 mins, i.e. 8 minutes). I think we can squeeze
this 8 minute-test-jobs into the build phase where we have to wait
anyway.

Also, we don't have to upload and download the build directory which is
a big plus (and can save some time too).

No recursion, just iterations using stack.

So far only the PrefixRankedTree -> NPDA version existed. This commit
adds a version that constructs an NFTA for RankedTree.

closes #203 - somehow it works if I broke string parser of automata it must have
been in a way tests didn't report it.

closes 227

closes 229

It seems we do not need to alter anything in the spec file.

By pulling as much images as we can from local cached copy we could
(hopefully) avoid the docker TOOMANYREQUESTS error.