Support more content types for the NSFWProtector

[Half-Shot]

In order to work around the limitations of tensorflow's models, this PR transcodes anything that might look like an image or video into something we can then pass through the model.

Tested and working.

The list of improvements are:

• Non-standard image types and video types are now thumbnailed.
• We cache the result of the last classification so repeat media are quickly deleted.
• Models have been updated.
• It's possible to specify the model you'd like to run with in the config.

[H-Shay]

@Half-Shot just curious about the status on this? I think we can probably merge it but I've lost context and want to make sure there's nothing I am forgetting.

[Half-Shot]

Hey @H-Shay, I'm mostly happy with this but I want to do more testing before we merge. I can take a look this afternoon.