[bug] faster implementation of NpzSNPSource._guess_filename

the previous implementation could take a very long time in some cases (especially on large datasets) owing to the fact that os.walk does not walk over the directory structure in lexicographic order by default.

in short, this meant that it could take a very long time to find the "matching" file.

this new implementation, which prevents seeking down directories that obviously don't match (e.g. the wrong scenario) allows zeroing in on the file being searched much more quickly