impr(store_volatile): collapse to a single ordered_set, LMDB-shaped

[shyba]

alternative to #879 -- instead of a helper indexer, change the main table to ordered set

What

Replaces current layout with a single ETS ordered_set, LMDB-shaped: {Path, {raw,V} | {link,T} | group} where list/3 is a prefix range scan with subtree-skip; group membership is no longer materialised.

Why

The bag-indexed design (PR #879) removed the expensive sets:add_element ordset-copy, but ets:bag insert scans the same-key chain for dedup — single-parent-heavy writes were also still slow.

A single ordered_set makes insert O(log N) unconditionally, removes the cross-table consistency complexity entirely, and matches LMDB/FS semantics (group is a marker, children are discovered by range scan). Tradeoff: point reads drop from O(1) hash to O(log N) tree; in practice the suite shows reads holding within noise.

How

• One ordered_set per store instance, keyed on path; values {raw,V} / {link,T} / group.
• list/3 is ets:next from <<Path, $/>>, accumulating the first segment of each descendant into a #{} map. After each immediate child it jumps past that subtree via <<Prefix, Name, $0>> (one byte past $/), so deep groups don't pay per-descendant cost.
• Root listing seeds from ets:first and skips the root marker — handles keys whose first byte sorts before $/ (e.g. base64url's -).
• write/3 / link/3 collapse to a single ets:insert/2; ancestors are walked once and each gets insert_new of group.
• make_group/3 is idempotent: leaves an existing group alone, converts a raw/link to group otherwise.
• Subtree overwrite uses ets:select_delete with a binary_part prefix guard — no recursion.
• list_path/3 returns {error, not_found} for an empty root so a chained store fallback isn't masked by {ok, []}.

No change to external semantics. read/3, type/3, resolve/3, link-of-link, group-of-links all preserved.

Measured impact

hb_store benchmark suite, scale=0.5 (50 000 records, 1 KiB values):

test	pre-index	bag-indexed (#879)	this PR
key write	2 231 /s	3 711 /s	519 k /s
key read	664k /s	637k /s	554 k /s
list write	9 961 /s	64 955 /s	291 k /s
list ops	523k /s	389k /s	210 k /s
message write	905 /s	1 363 /s	1 779 /s
message read	2 792 /s	4 351 /s	4 558 /s

Single-parent write goes from 1.66× (bag-indexed) to ~233× vs pre-index, eliminating the O(N²) tail. List ops regress from 523k/s to 210k/s without re-introducing O(N²) writes;

Tests

15 cases, all isolated via hb_test_utils:test_store/2:

• max_ttl_test — periodic reset.
• Overwrite invariants: overwrite_link_to_raw_test, overwrite_group_to_raw_test, overwrite_group_to_link_test, implicit_group_conversion_test.
• Listing semantics: list_test, list_dedup_test, list_with_link_test, list_root_test, list_root_includes_pre_slash_keys_test (regression: --prefixed keys), list_deep_subtree_test (subtree-skip), list_large_flat_group_test (10k siblings, dedup performance), prefixed_sibling_no_duplicate_test (interleaved sibling between a child and its subtree), sibling_with_zero_suffix_test ($0 jump-token boundary).
• empty_root_reports_not_found_test — store-chain fallback regression.