Fix: Load structs with large number of fields

[foreverallama]

Fixes #78 bug where structs with large number of fields could not be written in v7.3 format.

• MAT_HDF.jl: New method _write_field_reference!() to write field name attributes as a dataset and return HDF5 reference
• New read and write tests for struct_large.mat

Struct field names are written as attributes in v7.3 format. However, HDF5 has a maximum allowable attribute size of 64KB. When the number of fields is large, MATLAB instead writes a reference to a dataset, and writes the field names under #refs#.

The threshold for this write mechanism is when the total character length of all field names of the struct is >= 4096

[matthijscox]

This code should probably be added to more struct-like writing? MatlabClassObject, MatlabOpaque and MatlabStructArray perhaps?

[foreverallama]

Right, I forgot about MatlabStructArray, it should be added there. I don't think it needs to be added for the others but we can do it to be on the safer side

[foreverallama]

A couple of changes to support any type of struct data:

• Moved everything to a function _write_struct_fields() which invokes _write_field_reference() if required
• New constant struct_field_attr_matlab = "MATLAB_fields", replaced all instances of MATLAB_fields with that
• Support for MatlabStructArray, MatlabClassObject and scalar structs

The test is failing because for some reason the fields are going wildly out of order between save-load. Not sure what's going on here or how to resolve it. The data is intact though

[matthijscox]

I've made isapprox testing possible, and made sure it doesn't care about name ordering in MatlabStructArray isapprox.

[foreverallama]

Seems to be a quirk of HDF5 files. Keys are being written based on hash order of the Dict, but HDF5 stores/retrieves them lexicographically I guess