[FR] Update Lightning Flash integration to remove classes and default_classes usage

[ehofesmann]

Following #1858, the Lightning Flash integration needs to be updated to allow the reference to default_classes to be removed from the Lightning Flash docs discussed in this thread #1953. Note that since this issue there have been changes made to Lightning Flash where this is now only an issue for object detection datasets (and the only reference to default_classes in our docs is in the detection example).

These updates are primarily over in the Lightning Flash repo.

1. Specifically, removing the parsing of dataset.classes and dataset.default_classes here: https://github.com/Lightning-AI/lightning-flash/blob/ff92d4475cec09fc5f1dfcbcba753ce4bb353f09/flash/core/integrations/fiftyone/utils.py#L111

2. As well as adding a new argument allowing a user-provided labels list, which is still important so that train/test/val splits all share the same labels even if one of the splits does not include an example of a given class. Without this argument, the splits each have their labels parsed by this: https://github.com/Lightning-AI/lightning-flash/blob/ff92d4475cec09fc5f1dfcbcba753ce4bb353f09/flash/core/integrations/fiftyone/utils.py#L119

Specifically, I believe the from_fiftyone() method for detection data modules need to include this labels argument.