feat: save task parameters into cache and model checkpoint

[clement-pages]

This PR introduces two features:

• Save the task parameters into the cache. Indeed the values taken by certain task parameters can influence data preparation. Thus, generating a cache with particular values for these parameters, then reusing this cache with a new task, means that the parameters of this task may not be fully taken into account, and may distort certain results without raising the slightest error. This PR proposes to add a warning to inform the user that there are differences between the parameters of the current task and the one used to generate the cache.
• Save the task parameters into a model checkpoint. This makes it easier to resume training. For instance:

registry.load_database("/path/to/database.yml")
protocol = registry.get_protocol("AMI.SpeakerDiarization.only_words", preprocessors={"audio": FileFinder()})

model = Model.from_pretrained("/path/to/checkpoint.ckpt", protocol=protocol)