Remove role attribute from figure tags generated by filter_caption

Problem/Motivation

In reviewing #2994702: Allow editors to alter embed-specific metadata, as well as `data-align` and `data-caption`, @andrewmacpherson noticed:

The caption filter in core adds a role="group" attribute to the figure tag it creates. It would be more accessible to have that either be role="figure", or (even better) just removed entirely.

Proposed resolution

Let's do that. Kill the role attribute added to captioned figures!

Remaining tasks

Roll a patch, get tests written (or, really, existing tests changed as needed), and ensure this doesn't break backwards compatibility. Then commit it.

User interface changes

None.

API changes

None.

Data model changes

None.

Release notes snippet

TBD