Shared memory issue (#120)

[SCiarella]

This PR targets issue #120.

The test run (on Discord) with 128 MPI ranks fails because all ranks simultaneously load large (~1.2 GB) HDF5 opacity files into private RAM. With 5 molecules and 128 ranks, that's about 768 GB allocated on one node.

Every rank called self._spec_dict["xsecarr"][()] which reads the entire dataset into private RAM before moving it to shared memory. Fix: only the shared-memory root rank (rank 0 per node) loads the data from disk. Non-root ranks pass None to allocate_as_shared() and attach to shared memory without ever holding a private copy.

Also this line mem = GlobalCache()["xsec_in_memory"] or True is problematic because it always returns True, so it was impossible to ever disable in-memory loading.

I have also added some explicit close() for the HDF5 files.

---

With these modifications, the test does not fail anymore of oom issues, so I invite you to test this fix.

---

(Summary) Memory impact:

• Before: ~768 GB (128 ranks x 5 molecules x ~1.2 GB each)
• After: ~6 GB (1 copy per molecule in shared memory, shared across all ranks on the node)

[HannoSpreeuw]

• Before: ~768 GB (128 ranks x 5 molecules x ~1.2 GB each)

• After: ~6 GB (1 copy per molecule in shared memory, shared across all ranks on the node)

Wow, that's impressive.