Extract the module graphs through a staging directory

[traviswheeler]

Concurrent BLIMMP processes could not extract the module graphs at the same time. On a
9-genome Nextflow run, 4 of 8 tasks died before any analysis ran:

[FATAL] Failed to extract KEGG_Graphs_Generated_March26.zip. Tried ... and
/tmp/blimmp-cache-62226/KEGG_Graphs_Generated_March26.

Nextflow dispatches one BLIMMP task per genome, so the tasks landed on one node together
and all extracted into the same directory. _extract_module_graphs() extracted in place
and then mutated the result, removing __MACOSX and flattening the nested folder. One
process would delete or move files while another was still reading them. The OSError
that followed looked to the caller like an unwritable destination, so it tried the next
candidate and then reported that every candidate had failed.

Extraction now happens in a private staging directory that is renamed into position. A
rename onto a missing or empty directory is atomic on POSIX, so the first process to
finish wins and the others use its copy instead of failing.

Commits

The first commit adds the test without the fix, so CI records the failure before the fix
lands:

commit	Concurrent extraction step
3401ada test only	fails, 6 of 8 workers dead with the same FATAL
1efc959 fix	all workers pass; a bug in the test's own assertion kept it red
bc68eda assertion	passes, 8 workers, 340 graphs, no leftovers

Scope

0.1.4 makes this unreachable in the published image, because the graphs ship pre-extracted
and the extraction path is never entered. This matters for source and wheel installs, and
for any future image that stops pre-extracting. No release is needed to keep the container
behaving correctly.

The existing smoke test runs one process and stayed green through all of it, which is how
the bug shipped. tests/concurrency_test.py runs 8 workers against one directory and
checks that they all exit cleanly, agree on the result, and leave no staging directories
behind.