Lighter install

[leondz]

Machine learning code can come with a heavy install dependency chain, and garak is not free from this burden. We should make as much as possible of this optional, using the deferred loading mechanism.

Potential work includes:

• Remove generator plugins from default requirements, handle their non-presence with an error msg feature: disable optional imports by default #1475
• Make items with heavy dep chains, like transformers and pytorch, possible to avoid (with reduced functionality)
• Add a task to pre-cache required resources (from nltk, Hugging Face)
• Function to check dependencies required against the config of a run, and see if the install is already complete

Benefits include:

• Quicker garak install
• Lower space requirements by default
• Enables airgapped mode, with all data and dependencies loaded locally already