* Update demos/python_demos/formula_recognition_demo/README.md

Co-authored-by: Roman Donchenko <roman.donchenko@intel.com>

* upd readme

* use named constant for resolution setter

* Update demos/python_demos/formula_recognition_demo/formula_recognition_demo.py

Co-authored-by: Roman Donchenko <roman.donchenko@intel.com>

* move formula_rec_demo according to new layout

* regnetx-3.2gf: use the weights file that we download

Previously, it wasn't used in any way; the model constructor just downloaded
the weights anew.

* fastseg-*: use our ONNX converter instead of the one from the repo

The only differences in how they do the conversion compared to us is:

* they use the keep_initializers_as_inputs option, which doesn't do anything
  helpful (it just adds useless inputs to the ONNX model, which are stripped
  by MO);

* they use the verbose option, which dumps a lot of unnecessary information
  to stdout.

Using our own scri...