Add a `sampledata` module to mirror hvsampledata's API?

[maximlt]

hvsampledata has just been created and I would like to discuss whether we should integrate it further in hvPlot. This new package offers a simple API to load a dataset. Here penguins is a function that takes an optional engine (e.g. 'polars'), optional engine kwargs, and an optional lazy keyword.

from hvsampledata import penguins
df = penguins()

We can already use it to improve our examples:

```python
import hvplot.pandas
from hvsampledata import penguins

df = penguins()
df.hvplot()

I wonder whether we should add a sampledata top-level module that basically mirrors hvsampledata's API (i.e. from hvsampledata import penguins, ...) to make it even more accessible:

import hvplot.pandas

df = hvplot.sampledata.penguins()
df.hvplot()

As a developer, I quite like the idea that I could display a plot in 2 lines:

import hvplot.pandas
hvplot.sampledata.penguins().hvplot()

or

import hvplot.polars
hvplot.sampledata.penguins('polars', lazy=True).hvplot()

[maximlt]

A rough implementation of this module would look like the following:

# sampledata.py

try:
    import hvsampledata as _hvs
except ModuleNotFoundError:
    raise ImportError("Install hvsampledata to enable the sampledata module") from None

penguins = _hvs.penguins
...

__all__ = [
    'penguins',
    ...
]

The datasets exposed would only be the ones used on hvPlot's website.

We would not add hvsampledata as a dependency of hvPlot, neither in its pyproject.toml or its conda recipes. We could do that in the future based on our experience and user feedback. We could add it to an extra if that can help users getting started, that would be documented on the website.

[maximlt]

@holoviz/hvplot-dev what do you think about this proposal?

[maximlt]

The way hvsampledata is currently implemented means that calling hvsampledata.penguins() doesn't necessarily return a Pandas DataFrame. Instead, hvsampledata loops through a lookup table that links a data package to a format and its read function (e.g. pandas -> csv -> read_csv). If it finds the package is installed, it will import it and use its read function to load the dataset.

In order to make things more predictable, I guess it'd be best to pass the engine explicitly in the docs:

df = hvplot.sampledata.penguins("pandas")

[maximlt]

This suggestion wouldn't work if hvsampledata is not made a dependency of hvPlot or if the sampledata module isn't lazily loaded. Let's scratch that for now and revisit in the future if we feel the UX needs to be improved.