Temporarily rollback version of systran-storages

Larger context improvements (#448)

Fix error when the annotation is not a string

Copy the annotations into a standard Python dict

Support reading corpus from a JSON structure (#447)

Ignore "preprocess" parent model when resolving the models to average (#446)

Save the checksum of each model file in manifest.json (#444)

Include line number in Unicode decode error (#445)

Update pyonmttok to 1.37.1

Update pyonmttok to 1.37.0 (#443)

Define abstraction to allow sampling from other file formats (#441)

Add random option for larger contexts training (#442)

Update module systran-storages to latest (#440)

Improve message when a data file contains invalid Unicode characters (#439)

ref: ST-1516, allow hypothesis/reference normalization before scoring (#438)