Add array_transform to SolutionInfo to preserve compute-time transducer pose

Add an optional array_transform: ArrayTransform | None field to
openlifu.db.session.SolutionInfo capturing the transducer array-to-volume
transform matrix that was used when the corresponding Solution was
computed.

Motivation

Downstream (SlicerOpenLIFU) currently renders a solution's PNP / intensity
volumes by parenting them under the transducer's transform node, then snapping
the transducer to whichever virtual-fit or transducer-tracking result is
currently approved. This means:

1. If the currently-approved VF / TT changes between compute time and display
   time, the PNP moves to the new pose rather than staying where it was
   computed.
2. For a pre-solution (computed off a VF result), the VF approval may later
   be revoked; there is then no way to display the solution at its original
   pose because no transform survives.

The fix is to persist the exact matrix used at compute time on the
SolutionInfo provenance record. Consumers snap the transducer to that
matrix on display, so the PNP is invariant under later approval churn.

Details

• New field SolutionInfo.array_transform: ArrayTransform | None (default
  None for legacy round-trips).
• __post_init__ accepts a dict transparently so Session.from_dict
  decoding solutions entries produces the correct type.
• Round-trip tests added in tests/test_database.py.

Downstream

Tracked in SlicerOpenLIFU#622: "Pre-Solution PNP overlay renders at wrong
position when navigating to Sonication Planner". That issue's fix consumes
this new field.

Start with openlifu.db.session.SolutionInfo and inspect how Session.from_dict decodes solutions entries. Update tests/test_database.py for the optional array_transform field, including dict decoding and legacy None behavior; done means the value survives database round trips.