FEATURE: Support all bbcode tags in the rich editor - #107

Opening a post that used any of this plugin's tags ([color], [size],
[left], [list=a], etc.) in the rich text editor failed with "The rich
text editor doesn't support all features used in this post" and forced
the markdown editor.

Add a rich editor extension covering every tag the plugin supports:
inline styling as marks, alignment, [indent], [ot] and [edit] as block
nodes, and typed [list=X] lists. A tag value is only accepted when it
matches the charsets the cook sanitizer allows, so the editor can't show
styling the rendered post drops, and a value can't carry extra style
declarations into the editor's DOM.

Content the editor can't represent exactly is declined rather than
rewritten, leaving the post to the markdown editor with its source
intact: values outside those charsets, and a same-type tag nested in one
with a differing value, since a mark set holds one mark per type and
nested [size] percentages compound when cooked.

Cooking changes with it, without changing what a post renders as. Cooked
sepquotes carry a data-tag, so [ot] and [edit] are told apart
structurally instead of by their localized label, which no longer works
once a post is read under another locale. Typed lists get their own token
type and every item's content is wrapped in hidden paragraph tokens,
which render as nothing but let prosemirror-markdown's own list specs
parse them.

Track whether an inline bbcode open token was ours, so its close ends
the mark we opened rather than whichever one happened to be on top. The
block wraps already did this; the inline path popped unconditionally.

Build the cook sanitizer's allowlist and the editor's value checks from
one set of charsets, so the two can't drift. The shared module sits
under discourse-markdown/ because only that path is loaded into the
server-side cooking context.

Guard the parse state reads that assumed a node was open, and say in a
comment why declining a token fails the parse.

Assert that a declined post leaves the editor empty rather than that it
renders no span, which a silently dropped tag would also satisfy, and
drop the round trip cases that restate an exact round trip already
asserted in the extension test.