Prune unsupported global-level trace attributes

[alexcjohnson]

There are a bunch of attributes that we insert in the schema, and largely include in _fullData for all trace types. But many trace types don't support all of these attributes. As a simple workaround in @jonmmease's parcats PR, 273d2d0 (following @etpinard's suggestion #2963 (comment)) hides a the ones parcats doesn't use from the schema, but they will still show up in _fullData.

As I said in #2963 (comment):

There's a module categories: ['noOpacity'] which prevents coercing opacity (so it's not in fullData, which is important for the editor) but does not take it out of the schema (and anyway only handles opacity). @etpinard 's solution would fix the schema but not fullData. Sounds like what we really want is something like:

module.omitGlobalAttrs: ['selectedpoints', 'hoverlabel', 'opacity', 'ids', 'customdata']

that would do both.

There are probably quite a few other traces that omit some of these - in fact, probably everything that sets traceOut._length = null has no use for selectedpoints, ids, and customdata...

And @etpinard responded #2963 (comment):

Good point here. But I think for this PR my suggestion in #2963 (comment) will be good enough. The react-chart-editor (aka RCE) doesn't know how to handle traces with dimensions yet, so having a few extra keys in fullData won't affect anyone at this stage. But yeah, once merged we should open a new issue (similar to #2834 and #2908) about this.

So here's that issue :)

[alexcjohnson]

There's a related issue with the way we merge base and module attributes in the schema: sometimes extendDeepAll goes too far, as in:

> Plotly.PlotSchema.get().traces.scatterpolar.attributes.hoverinfo.flags
<- ["r", "theta", "text", "name", "name"]

where the second "name" came from the base ["x", "y", "z", "text", "name"] getting added to a length-4 override in the module. We should really do a more manual crawl, only adding base attributes if they haven't been overridden in the module.

Anyway a lot of modules that have no "z" inappropriately inherit it as well, this should really perhaps always be handled at the module level.

[etpinard]

After @antoinerg 's #3158

• hoverinfo flags are now extended properly in trace attribute module
• hoverlabel is now omitted in attributes and not coerced for traces that don't support it

Now, me and @alexcjohnson disagree on what to do with ids and customdata. We'll leave this issue open until we reach a consensus.

[antoinerg]

After @antoinerg's #3158

• transforms was removed from carpet, cone, contourcarpet, mesh3d, heatmap, parcoords, pointcloud, sankey, streamtube, surface and table

[etpinard]

• Traces w/o the showLegend category should not have showlegend, legendgroup in their schema