fix(trainer-rank): credit staged adapter gradients and shared captures in graph admission

[bradhilton]

Agent: Dalinar

#925's graph-placement admission over-prices the first gradient wave of a slot by charging the same memory twice:

1. Adapter gradients: it reserves 3× the slot's parameter bytes as gradient staging (current, staged and replacement), and also prices fix(trainer-rank): account for pending adapter gradients during recompute #1002's pending adapter gradients inside the unit's workspace. Those pending gradients are the staged gradients, the same bytes.
2. LoRA version snapshot: the snapshot is charged once per unit or split child, but every unit with the same slot and staleness shares one capture.

This change credits the staged adapter gradients against the unit's workspace, so the full 3× staging reserve stays and rollback, the multi-forward prepare and the in-flight margin remain covered. It also charges each shared snapshot once. Nothing else in admission changes: CP no-release credit, custom heads, checkpoint sharing and learned split floors are untouched.

On a Qwen-like CPU fixture (40 layers, 920 MiB slot):

Case	Admitted on main	Admitted after
first wave, 1,024 rows	5,865 MiB	5,101 MiB
first wave, 4,096 rows	6,431 MiB	6,135 MiB
later waves	unchanged	unchanged
split, two same-slot children	6,244 MiB	4,716 MiB

Testing: new admission tests fail on main and pass here. 2,257 trainer-rank CPU tests pass. GPU CI runs on this PR.

Limits:

• The arithmetic is CPU-derived from the code's allocation lifetimes, not a GPU measurement, and the fixture numbers are not a GPU qualification. The credit equals fix(trainer-rank): account for pending adapter gradients during recompute #1002's net pending adapter-gradient contribution, not the full staged tensor bytes.
• Admission still over-prices in places. The replacement-gradient share never coincides with a backward, but removing it needs a measured in-flight gradient margin, and warm profiles may already include the snapshot. Both are left for a GPU-logged run.

🤖 Generated with Claude Code