Better metadata loading for toyfiles

[hammannr]

This pull request changes how metadata is loaded so that it's more usefull. Instead of just returning a dict for each file, we now add an entry "metadata" (plus an entry for each of the array_metadatas) to the returned results dict that is a structured array and has the same dimensions as the results.

This way you can easily relate information from the metadata to each toy experiment.

In addition, I added a short docstring and raise an fnf error in case the patterns don't match any file.

[kdund]

Hi, @hammannr how does this change the size of files?

[hammannr]

@kdund the files on disk don't change, this just makes the already stored metadata more accessible.
The memory usage of course increases when loading the metadata and this depends on the number of array names and number of fields in the metadata etc.
To get an idea, I get the following deep sizes of the returned dictionary (using pympler.asizeof) for a random example with 3 array names and 90_000 toys :

• w/o metadata: 25 MB
• w/ metadata: 33 MB

[dachengx]

Thanks @hammannr . I fixed a small bug when numpy_array_names is not None. Let's go with it.