Share the NPZ object serialization of the Web API between DataLab and DataLab-Kernel #90

Problem

The NPZ format of the DataLab Web API (in-memory archive of NumPy arrays and JSON metadata, used to transfer signals and images) is implemented twice, by copies that diverge:

• DataLab, datalab/webapi/serialization.py (server side);
• DataLab-Kernel, datalab_kernel/serialization_npz.py (client side).

Each fix must be done twice: non-uniform image coordinates were added to DataLab (DataLab-Platform/DataLab@34ad31c) and only later to the Kernel (DataLab-Platform/DataLab-Kernel#4); the Kernel still cannot transfer object metadata holding NumPy arrays nested in lists (e.g. ROI-related results), which DataLab handles.

Proposal

Both depend on Sigima, which defines the objects: provide the implementation once, in sigima.io (serialize_object_to_npz, deserialize_object_from_npz), taken from DataLab (the most complete): labels, units, uncertainties, uniform or non-uniform image coordinates, object metadata (NumPy arrays and scalars converted, geometry coords restored), checks of invalid archives, compatibility with archives written by earlier versions, no pickling. The format is documented next to the code.

DataLab and DataLab-Kernel then use it instead of their own copies.