Resolve model engine names through the shared registry (item 5, step 2)

[isayev]

Non-breaking. available_models() returns the same six names in the same order,
and _adapters was private. Suite: 1756 passed, 1 skipped, 70 deselected.

Three lists holding different sets

list	contents
ModelFactory._adapters	2 — ANI2X, ANI2XT only
ModelFactory.available_models()	6 — a hand-written literal, derived from nothing
ENGINE_INFO (cli/commands/models.py)	6 — AIMNET, three aimnet variants, two ANI

_adapters held only engines built from a local adapter class. The four
aimnet entries are not adapters at all — they are names resolved through the
aimnet registry. So one list was keyed by "has a local adapter" and two by
"user-facing name", and nothing connected them; they agreed by hand.

This is the finding recorded in #162's body, and it is why step 2 was the larger
half.

What changed

_engines is one registry keyed by user-facing name, with the adapter class as
the value where there is one and None where the name resolves through the
aimnet registry. available_models() derives from it instead of restating it —
adding an engine and forgetting that literal used to leave it working but
undiscoverable.

The import-time assert set(_adapters) == set(BUILTIN_ANI_MODELS) survives and
is now derived from the registry, so the two cannot drift. It moved to module
scope because a comprehension inside a class body cannot see that class's own
attributes — a Python scoping rule, found by it raising NameError on import,
not a style preference.

ENGINE_INFO is deliberately still separate

Moving it into info= belongs with the CLI's changes rather than this factory's.
It is the remaining piece of item 5, and it is named here so the third list is
not mistaken for handled.