error out with a readable message in chwewckpoint merger for incompatible tensor shapes (ie when trying to merge SD1.5 with SD2.0)

maximum of 2 models loaded at once. delete unneeded model before next
step. fix 'teritary' -> 'tertiary'. gracefully fail when "add
difference" is selected without a tertiary model

rework #5012 to also work for pictures dragged into the prompt and also add Clip skip + ENSD to parameters

There is no need to re-hash the input image each iteration of the loop.
This also reverts PR #4026 as it was determined the cache hits it avoids
were actually valid.

This changes the extras image cache into a Least-Recently-Used
cache. This allows more experimentation with different upscalers
without missing the cache.

Max cache size is increased to 5 and is cleared on source image
update.

Bit of a refactor to the image cache to make it easier to extend.
Also takes into account the entire image instead of just a cropped portion.

Face restoration can look much better if ran after upscaling, as it
allows the restoration to fix upscaling artifacts. This patch adds
an option to choose which order to run upscaling/face fixing in.

Update serialization on the batch images endpoint

In batch mode run_extras tries to preserve the original file name of the
images. The problem is that this makes no sense since the user only gets
a list of images in the UI, trying to manually save them shows that this
images have random temp names. Also, trying to keep "orig_name" in the
API is a hassle that adds complexity to the consuming UI since the
client has to use (or emulate) an input (type=file) element in a form.
Using the normal file name not only doesn't change the output and
functionality in the original UI but also helps keep the API simple.

improve performance of 3-way merge on machines with not enough ram, by only accessing two of the models at a time

remove sigmoid and inverse sigmoid because they just did the same thing as weighed sum only with changed multiplier