Dependencies on Seaborn and TQDM are not implied by the installation of Jupyter Scatter

[hamelin]

When installing a package, it is customary to expect all its external dependencies to be pulled in implicitly. However, the dependencies on seaborn and tqdm are not. For instance, I create the following virtual environment, on a Bourne-compatible shell, and start Jupyter Lab:

python -m venv .venv
. .venv/bin/activate
pip install jupyterlab jupyter-scatter
jupyter lab

I then create a new notebook, and try to run

from jscatter import Scatter

The first time, it raises an exception saying that package seaborn cannot be imported. I pip-install Seaborn, then restart the kernel and retry. This time, the same import cell raises that package tqdm cannot be imported.

The solution would be to add these two packages to the formal dependencies of jupyter-scatter.

[hamelin]

So I've been trying to get a dev environment to fix this myself, but ran across that Seaborn and TQDM are both installed when using the all package option. These dependencies are thus themselves only required when using certain features of the scatter plot. However, the import of jscatter imports them in turn, regardless of what features are actually used.

I see thus two paths to fixing this:

1. Carry on with the labels and annotations package options, but guard against import failure of these dependencies (including that of the HDBSCAN package). Then we set up further guards everywhere in the code where annotations and labels come into play.
2. Make these packages all parts of the standard install. As far as I can tell, they don't carry such a specific dependency tail of their own as to make their deployment uselessly onerous. We may keep the install options as NO-OP for the time being, and deprecate them on some timeline, but unless the install base is already very large, discarding them should not carry that bad an impact.

I am strongly in favor of option 2. Many old data scientific Pythonistas naturally use Seaborn for this or that; TQDM is also a common idiom to many coders. I am biased with respect to HDBSCAN (I work directly with its authors), but it also consists of an inexpensive, short-tail dependency. It would free Jupyter Scatter development from the burden of keeping in mind that labels or annotations might not be available to the user. If one were to go with option 1, I would surmise that one would need the guards I mentioned peppered through many parts of the code base, which would likely be annoying.

I leave the decision to @flekschas . If any of the former options fit the bill for you, I will commit the time to implementing the fix.

[joshdunnlime]

I noticed the seaborn issue as well (TDQM was probably already install for me via another package).

[flekschas]

Arghh sorry about that. I know that Python loves to ship everything but that just eventually slows down install and compile time. So, I prefer Option 1. if you have spare cycles, please propose a PR. I'm happy to get it merged