NeoMatlabIO: round-trip quantity, array and nested annotations

[adityasingh2400]

When NeoMatlabIO writes an object, _get_matlab_value flattens the annotations dict for MATLAB. A quantity is split into a plain magnitude plus a companion <key>_units field, exactly the way quantity attributes such as t_start are stored. A nested mapping becomes a nested struct. None becomes the sentinel string Py_None, because MATLAB has no equivalent and scipy drops a real None.

The read side undid none of that. The elif attrtype == dict branch copied every field of the struct straight into the annotations dict, so a quantity came back as {'yop': array([3., 4., 5.]), 'yop_units': 'ms'} instead of [3, 4, 5] * pq.ms, and a nested mapping came back as a raw scipy.io.matlab.mat_struct object. That is what issue #852 reports.

The same branch has a worse problem that the issue does not mention. It tested the sentinel with if value == PY_NONE, and on an array-valued annotation that comparison returns an array, which is not usable as a condition. So ValueError: The truth value of an array with more than one element is ambiguous came out of read_block for any file holding an array annotation, quantity or not. The file was written without complaint and could then never be read back.

Decoding now mirrors the encoding, in a new create_dict_from_struct alongside the existing create_ob_from_struct. A <key>_units field is folded back into the value it belongs to and is not emitted as a key of its own, but only when the field it names is actually present, so an annotation genuinely called foo_units still round-trips on its own. A nested struct is decoded recursively. The sentinel is tested only on values that are strings. One line changed on the write side too: None inside a nested annotation dict was silently dropped, because the guard that names the annotations attribute did not survive the recursion into the nested mapping, and annotations are the only dict-typed attribute any Neo class declares, which I checked across class_by_name.

Verified with a write then read round trip on a temporary file, which needs no downloaded data. The new parametrized test covers a string, an int, None, a plain array, a quantity array, a quantity scalar, a flat dict and a doubly nested dict holding quantities, and asserts no companion field leaks into the annotations. The annotation the issue asks for, yop=[3, 4, 5] * pq.ms, is also added to test_write_read_single_spike as suggested. Against 35cbce7, neo/test/iotest/test_neomatlabio.py is 6 failed 11 passed before the source change and 17 passed after, and neo/test/coretest stays at 637 passed 0 failed.

One thing I want to flag rather than bury: a quantity and its units are two separate fields in the .mat file, so an annotation dict that genuinely contains both a and a_units is indistinguishable on disk from a single quantity annotation a, and now reads back as the latter. That ambiguity is inherent to the storage format the write side already used, and it is the same convention Neo applies to object attributes, but it is a behaviour change for that one case.

Fixes #852