Add support for online drift detectors

[ascillitoe]

This PR adds support for online drift detectors to Alibi Detect runtime.

Overview

This involves two primary changes:

1. Online drift detectors are run on one instance at a time, with the time-step self.t updated each time. If a request contains a batch of instances, the payload must be unpacked and passed to the detector one instance at a time.
2. Online drift detectors have state, which must be periodically saved for checkpointing. State is saved every state_save_freq time-steps, where state_save_freq is a new configurable parameter added to AlibiDetectSettings.

Limitations

Parallel inference is NOT currently supported for online detectors (i.e. parallel_workers=0 is required). This would require careful thought wrt to how state is managed/shared across parallel workers, in addition to considering the implications of distributing non-iid samples between workers.

TODO's

• Documentation? - should be auto-completed pending Fix settings reference in runtime docs #1109
• Linting

[ascillitoe]

Wrap X in a list (of length 1) so that we don't unpack into single instances when running in batch mode (i.e. offline drift detectors or outlier detectors). pred is then returned as a list of length 1.

This slightly circuitous logic is intended to avoid having two duplicate try-except blocks for the two (offline vs online) use cases...

[adriangonz]

This looks great @ascillitoe! 🚀

I've added a minor comment around the state save logic and check, but everything else should be good to go once the linter and tests pass!