Make convert_ldm_clip_checkpoint more flexible about maxtoken

[ppbrown]

Make the code examine input model shape to infer actual max token len,
rather than forcing CLIP default of 77.
Update feature_extractor to use AutoImageProcessor, because old code use caused problems.

@yiyixuxu and @asomoza

Fixes #11285

I tested the code by running the following script, and then generating an image from the resulting diffusers-format model.

# run in "scripts" dir
MDIR=~/models

HF=opendiffusionai/xllsd-alpha1
SRCMDL=$MDIR/xllsd
DESTCHKP=$MDIR/xllsd.safetensors
MDL=$MDIR/xllsd-new
if ! test -d $SRCMDL ; then
	mkdir -p $SRCMDL
	hf download --local-dir $SRCMDL $HF
fi

if test -f $DESTCHKP; then
	echo $DESTCHKP already exists
else
python convert_diffusers_to_original_stable_diffusion.py \
--use_safetensors \
--model_path $SRCMDL \
--checkpoint_path $DESTCHKP
fi

python convert_original_stable_diffusion_to_diffusers.py \
 --from_safetensors --to_safetensors \
 --dump_path $MDL \
 --checkpoint_path $DESTCHKP