Generic Pylaia worker

While the training worker is being implemented in #8 (closed) , we can already imagine the Pylaia generic worker.

This worker's user configuration parameters will at least include:

• use_language_model, bool, defaults to False, whether a language model is required (will fail if looking for one and not finding)
• from_right_to_left, bool, defaults to False, use Right-to-left orientation
• extraction_mode, enum, (see current param)
• batch_size, int, defaults to 2
• line_element_type, str
• line_worker_version_id, str
• scale_x, float
• scale_y_top, float
• scale_y_bottom, float
• color_mode (old image_convert), enum
• lm_weight, float

The current models have to be ported in the right format via the CLI, an archive containing

• model
• syms.txt
• weights.ckpt
• language_model.arpa.gz (optional)
• lexicon.txt (optional)
• tokens.txt (optional)

We need to do something similar to U-FCN generic worker:

• load the model version configuration using self.model_configuration
• load the model and optional language model in the right folder
  • if "model" is in config, keep the current behavior
  • else, find the path to the model using self.find_model_directory() and retrieve what's available there (language model if self.config[use_language_model] is True)

Just like U-FCN generic worker, we need a new CI job that publishes the model on https://arkindex.teklia.com. There is no need to rename the folder though in this case but we need the --use-parent-folder option to publish the whole folder instead of only the model binary file.