b+tree: index a node's entries through one accessor, and write up the occupancy work

[psiha]

Two commits.

1 — one entry accessor. Seventy-odd sites reached into node.keys[ i ] directly, which states in each of them that entry i of a node lives at offset i of its key array. That is true today and it is the only reason the tree cannot yet leave a gap at the front of a node.

key_at( node, i ) is that statement, made once. It is still exactly node.keys[ i ] — no behavioural change, same codegen — but there is now a single place answering "where does a node's first live entry sit", which is what a devector node needs and what a map's parallel value array needs.

The raw-array search helpers keep taking a bare Key const keys[]: their caller has already resolved the base, and they must stay callable on a span that is not a node at all (the bulk merge path passes source keys straight in).

2 — doc/b+tree_occupancy_and_variants.md. What fill a b+tree actually reaches, what the classical variants do about it, and what shipping systems actually do — read at source rather than taken from secondary literature:

• No true B* among the eight engines checked. SQLite's balance_nonroot redistributes over up to three siblings n→n+1 and its delete-side floor is ~1/3, the opposite of B*; Apple's HFS docs say "B*-tree" but InsertLevel rotates into the left sibling only and then splits 1→2. PostgreSQL, InnoDB, LMDB, Berkeley DB and WiredTiger all split 1→2 with no sibling redistribution at all.
• The naming. Comer 1979 opens with "Perhaps the most misused term in B-tree literature is B*-tree": Knuth's B* is 2/3 occupancy maintained by local redistribution and a two-into-three split. What b+tree: relieve an overflowing leaf into a sibling before splitting #157 implements is the local redistribution without the floor and without the 2→3 split.
• Why not go further. Three independent reasons, including that min_values = ceil( max / 2 ) is not a constant that can be raised (see b+tree: state the bounds a node actually has #155) and that the one careful measurement of B* puts it at −15% space for +31% cycles, beaten on both axes by deferred compaction.
• Which of LeanStore's node-level techniques transfer to a container with no buffer manager and no synchronisation — and which plainly do not.

Stacked on #157.