Improve and clean up the visualisers

[kmaziarz]

The visualisers were recently improved in #10, but several confusing aspects still remain. This issue is meant for tracking all these quirks while they are being cleaned up.

• In the pictures of the partial molecules produced in HTML mode, only some of the atoms are displayed with IDs. However, atoms without IDs can still later appear in edge prediction steps, at which point it's not possible to match the IDs shown in the table to atoms in the picture.
• The atom/motif choice step omits the "no more nodes" class, which can be especially confusing at the last step when this is the ground-truth class.
• Steps are not numbered consistently (depending on the mode, either 0-based or 1-based), and the first node prediction step is not numbered at all. Moreover, the division of the different choices into steps is centered around edge/attachment decisions, with node decisions being glued to one of these, and this can sometimes lead to confusing outputs. In particular, in the "decode from samples" mode, in some cases the first two node selection steps are shown before the "Step 1" heading is rendered. It would be clearer to just show one step for every node/edge/attachment decision. One complication is that some of these decisions are skipped in the decoder if there is only one choice to be made (e.g. choosing an attachment point in a benzene ring); ideally, the visualisers would explicitly call it out and label it as "skipped".