Expose a uniform API at the highest level for models

[theabhirath]

This PR does a bunch of stuff:

1. Proposes to refine the API at the highest level of models to only expose options that make sense for the model to load pretrained weights (this means removing configurations like drop rates).
2. Completes the migration towards having inchannels and nclasses as uniform options for all models, helping along Add inchannels, imsize, nclasses as kwargs for all constructors #176 - only imsize is left now.
3. For the purpose of general code cleanliness, defines a ton of type annotations that weren't there before - these will help along with the documentation (once refactored) to help the users figure out more cryptic errors that may occur when dealing with lower level model APIs.
4. Does away with default model configurations at the highest level. I can restore this, but this behaviour is rather ambiguous - what's to make a ResNet 50 be more suited for a default configuration than a ResNet 18, or a 101?
5. Throws in a small refactor of invertedresidual as prep for an EfficientNetv2 PR that will be landing shortly.
6. Uses create_classifier in most cases where it can be for brevity
7. Adds compat entries for packages introduced in Overhaul of ResNet API #174 (closes CompatHelper: add new compat entry for CUDA at version 3, (keep existing compat) #191, closes CompatHelper: add new compat entry for ChainRulesCore at version 1, (keep existing compat) #192, closes CompatHelper: add new compat entry for PartialFunctions at version 1, (keep existing compat) #193, closes CompatHelper: add new compat entry for NNlibCUDA at version 0.2, (keep existing compat) #194). Also closes CompatHelper: bump compat for Functors to 0.3, (keep existing compat) #180 because why not

[theabhirath]

Pending CI, this should be good to go!