chebifier has many (too many) dependencies

[sfluegel05]

Since we are using models from different libraries, we have to install them as well as their dependencies. For chebifier==1.1.0 i got

Installing collected packages: pytz, mpmath, lark-parser, distlib, deepsmiles, antlr4-python3-runtime, urllib3, tzdata, typing-extensions, threadpoolctl, sympy, soupsieve, six, setuptools, selfies, safetensors, regex, pyyaml, pygments, propcache, pluggy, platformdirs, Pillow, packaging, numpy, nodeenv, networkx, multidict, MarkupSafe, lxml, joblib, iniconfig, importlib_resources, idna, identify, graphviz, fsspec, frozenlist, filelock, fastobo, docstring-parser, dill, colorama, charset_normalizer, cfgv, certifi, attrs, aiohappyeyeballs, yarl, virtualenv, typeshed-client, tqdm, scipy, requests, rdkit, python-dateutil, pytest, pbr, omegaconf, multiprocess, lightning-utilities, jsonargparse, jinja2, click, beautifulsoup4, aiosignal, torch, scikit-learn, pysmiles, pre-commit, pandas, huggingface-hub, aiohttp, torchmetrics, tokenizers, iterative-stratification, gavel, transformers, pytorch-lightning, chemlog, lightning, chebai, chebifier

which took a bit longer than probably necessary. I would suggest doing two things:

1. Check which dependencies are actually needed to run the ensemble (I am pretty sure torchmetrics for instance is not one of them, neither is graphviz), where to unnecessary ones come from and how we have to rewire the installation to keep them out
2. Make at least some models optional - this is already the case for chebai-graph, which is not installed by default

[cthoyt]

as a first time user, I can say that it's more of a problem for me to pip install this software and not have the example on the README work than having a lot of dependencies. It's better if everything for minimal downstream usage is default dependencies. Training dependencies can be optional