BUG: fix ParamFileStore for on-demand frontend imports

[chrishavlin]

Currently, the Dataset pickling is broken between python sessions when using the yt config option store_parameter_files = true. For example, given two scripts, create_a_pickle.py and load_a_pickle.py:

# create_a_pickle.py
import yt
import pickle

if __name__ == '__main__':

    fname = 'IsolatedGalaxy/galaxy0030/galaxy0030'
    ds = yt.load(fname)

    with open('yt_ds.pickle', 'wb') as handle:
        pickle.dump(ds, handle, protocol=pickle.HIGHEST_PROTOCOL)

# load_a_pickle.py
import pickle

if __name__ == '__main__':

    with open('yt_ds.pickle', 'rb') as handle:
        ds = pickle.load(handle)
   _ =  ds.index

The following fails on the attempt to load:

$ python create_a_pickle.py
$ python load_a_pickle.py
Traceback (most recent call last):
  File "/home/chavlin/src/yt_general/performance/yt_perf_testing/miscelleaneous_tests/dask_pickles_etc/pickle_from_file.py", line 5, in <module>
    ds = pickle.load(handle)
  File "/home/chavlin/src/yt_general/yt/yt/data_objects/static_output.py", line 2053, in _reconstruct_ds
    ds = datasets.get_ds_hash(*args)
  File "/home/chavlin/src/yt_general/yt/yt/utilities/parameter_file_storage.py", line 91, in get_ds_hash
    return self._convert_ds(self._records[hash])
  File "/home/chavlin/src/yt_general/yt/yt/utilities/parameter_file_storage.py", line 117, in _convert_ds
    raise UnknownDatasetType(class_name)
yt.utilities.parameter_file_storage.UnknownDatasetType: EnzoDataset

(I also ran into this problem when experimenting with some dask functionality...).

The issue comes how yt now imports frontends on demand (output_type_registry is reset between sessions).

[chrishavlin]

Setting to draft for now because I want to see if I can write a test with some pytest monkeypatching.

Also note that the fix here is not backwards compatible -- IMO it's fine because I don't think many folks are using this functionality and it's not meant for long-term dataset storage anyway.

[chrishavlin]

this is the potentially backwards incompatible part: when using the on-disk store, this new field will be a new column in the csv. so if an on-disk store already exists, there will be an error due to the different number of columns and the user would need to remove the old on-disk store.

[chrishavlin]

assuming tests pass this should be ready for review

[chrishavlin]

oh, those are real failures. will take another go at those pytest fixtures....

[neutrinoceros]

Any reason not to put this import at the top level ? importlib should be "free" to import anyway.

[neutrinoceros]

Not 100% sure this is equivalent (I don't know how many "." are expected), but I'm willing to bet there's a clearer way to express this change in file extension.

Suggested change

	fe_api = ".".join(module.split(".")[0:3]) + ".api"
	fe_api = module.rpartition(".")[0] + ".api"

[neutrinoceros]

Fixtures that are used as setup/teardown but don't need an accessible name should be declared with pytest.mark.usefixture