b+tree: relieve an overflowing leaf into a sibling before splitting

[psiha]

A split-on-full b+tree does not sit at its minimum fill — but it does not sit near full either. Measured by the accompanying characterisation test, 8M uint32 keys:

pattern	before	after	leaf bytes/key
random, 512-byte nodes	69.5%	87.1%	5.94 → 4.74
sequential (either direction)	50.0%	99.2%	8.26 → 4.16
random, 4096-byte nodes	72.8%	89.9%	5.52 → 4.46
merge, interleaved (4096)	67.6%	88.2%	5.94 → 4.55
bulk / appended merge	100% / 99.9%	unchanged	—

Random insertion settles at Yao's ln 2, not at 50%; sequential insertion settles at exactly 50%, which is where the waste actually is. The bulk paths are already at capacity and stay there.

So before splitting, an overflowing leaf now hands values to a same-parent sibling that still has room. This is Comer's local redistribution — and deliberately not a B*: no 2/3 floor, no 2-into-3 split, minimum fill untouched, no merge code changed. Structurally it is the exact dual of handle_underflow's borrow branches and uses the same idioms:

• sibling existence resolved from parent_child_idx, never the level links — those cross parents;
• the separator follows the values;
• half the room found is taken, not all of it — a sibling emptied of slack merely moves the next split one node over, and a sibling left full would have to be split by the very insertion being relieved.

Leaves only, deliberately. A node's children carry a back-index into their parent, so relocating an inner node's children re-indexes and dirties every one of them — more than the split it would save.

Confirmed on the resident node pool, not just leaf bytes (hence nodes_used()/nodes_reserved()): pool bytes/key 6.09 → 4.87 random, 8.53 → 4.30 sequential.

Cost. Inserts are neutral-to-faster throughout. The one measured regression is find() in the 512-byte configuration, 1.08×, where intra-node search is a linear scan and a fuller leaf is a longer one; at 4096 bytes, where that search is binary, the same measurement is 0.95×. That is a property of occupancy, not of this policy — a bulk-built tree pays it too. The structural cost is that relieving leaves a node nearly full so it overflows again sooner; each event is individually cheaper, and to_move = room / 2 is the knob.

On by default; PSI_VM_BT_REDISTRIBUTE_ON_OVERFLOW=0 restores plain split-on-full.

Background — measured occupancy, the B*/Comer naming history, an eight-system prior-art table, and why this is not becoming a B* — is in doc/b+tree_occupancy_and_variants.md (#158).

Stacked on #156.