Switch to categorical handling when color_key specified

[jlstevens]

This issue is one of a number that relate to datashader support of timeseries. In particular, this issue relates to https://github.com/holoviz/hvplot/issues/939 which is also about switching to categorical handling.

>>> hv.__version__, bokeh.__version__, ds.__version__, panel.__version__, hvplot.__version__
('1.15.1', '2.4.3', '0.14.2', '0.13.1', '0.8.1')

>>> timestamps = [pd.Timestamp('2022-04-01 0{0}:00:00-0000'.format(hour)) for hour in range(10)] 
>>> df = pd.DataFrame({'timestamp':timestamps, 'SPY':np.random.rand(10), 'NDAQ':np.random.rand(10)})
>>> melted = pd.melt(df, id_vars=['timestamp'], var_name='stock', value_name='high')
>>> melted['bond_indicator'] = melted['stock'].isin(['SPY'])

There are multiple problems illustrated here due to color_key not forcing categorical handling:

1. There is a colorbar when there shouldn't be one as this should be categorical (only two colors, red and green)
2. There is a legend but it isn't actually working properly (it is blank). This may just be bokeh having difficult showing legends for images.
3. There is a mixture of red and green for all the lines! This is because instead of being treated as a categorical, the curves are being antialiased (line_width=1) which gives you floats, with a stronger value in the middle of the line and lower values near the antialiased, fuzzy edges.

By using line_width=0 it is possible to see that things weren't actually called correctly (only red lines):

Here is what does currently work:

However, you need to use aggregator=ds.by('bond_indicator') when the intuitive thing to do is color='bound_indicator'. The latter almost works (but not quite due to the lack of categorical handling!)

Now the 'red' color is wrong! Turns out this is due to histogram equalization in this non-categorical (float) case:

What I believe is that both color='bond_indicator' and color_key=["#FF0000", "#00FF00"] should force categorical: the former as the dtype is boolean and categorical should always be used when color_key is specified. Issue https://github.com/holoviz/hvplot/issues/941 has some suggestions for what to do when these options clash...

[jbednar]

I agree that both color and color_key should indicate categorical unless otherwise overridden. I'm not sure what should force non-categorical processing.