Consider using a more accurate nsfwjs model

The default one seems slightly inaccurate in some cases, for example the default model identifies this image of my element web call stack as a high chance of being porn
| Default | MobileNetV2Mid |
|--------|--------|
|  |  |

In particular, i suggest changing to https://github.com/infinitered/nsfwjs?tab=readme-ov-file#load-the-model "MobileNetV2Mid", the InceptionV3 model might be interesting too, but not sure how it would affect size

Start by locating where mjolnir loads or configures the nsfwjs model, then read nsfwjs's model-loading guidance for MobileNetV2Mid and InceptionV3. Compare the current default with the proposed model for classification accuracy and size; done means a justified default choice is implemented and verified against the reported false positive.

Written by the indexing model from the issue text.