impr(store_volatile): index group children in a separate ETS table

[shyba]

What

Replaces the volatile store's in-band child set with a dedicated
children index table, adding an overwrite fast path to write/3 and
make_link/3.

Why

hb_store_volatile tracked group membership by storing a sets:set()
of children inside the main table's {group, Set} row. Every write
walked the parent chain and called sets:add_element/2, which copies
the ordset.

How

• Second ETS table per store instance, bag-typed, keyed by parent
  path, each row {Parent, ChildName}. Bag semantics collapse
  duplicate rows automatically.
• Group markers in the main table simplify to {group, true}. All
  membership lives in the index table.
• list/2 on a group path becomes a single ets:lookup/2 on the
  index — O(family-size), no scan.
• write/3 and make_link/3 branch once on the existing entry:
  • absent → run parent-index maintenance as before
  • existing {group, true} → delete the index rows for that path
    before overwriting, so the new raw/link value doesn't leave
    stale children visible on later recreate
  • existing raw/link → skip parent-index maintenance (already
    indexed when the key was first created)
• reset/1 clears both tables.

No change to the store's external semantics. list/2, type/2,
make_group/2, read/2, resolve/2 behavior is preserved.

Measured impact

Direct writes (1 KiB value, single parent):

N	before	after
5 000	26k ops/s	1.12M ops/s
20 000	6.9k ops/s	995k ops/s
50 000	2.2k ops/s	1.13M ops/s

End-to-end hb_cache:write on volatile, 200 1 KiB committed
messages, 10-run median:

• before: ~590 msgs/s
• after: ~1714 msgs/s

For comparison, hb_store_lmdb on the same workload: ~1800 msgs/s.

Tests

New coverage:

• list_test, list_dedup_test, list_with_link_test — group
  listing and dedup invariants.
• overwrite_link_to_raw_test — raw value replacing a link.
• overwrite_group_to_raw_test — raw value replacing a group
  marker; asserts index rows for the path are gone before recreate.
• overwrite_group_to_link_test — link replacing a group marker;
  same cleanup invariant, verified through list/2 via link resolution.

[shyba]

Back to draft for 2 reasons:

• there is a better way to do it also using two tables, like this one
• another PR will show another way of redesigning it entirely, so it goes back to 1 table

[shyba]

Replaced by #887