Implement the agreed serialized-magnet accessor convention

[GamelinAl]

A serialized family (N magnets on one power supply) has a single degree of freedom, the PS current I. The accessors now follow that in design and live:

• family.hardware returns I (was N x I) and sets I.
• family.strength returns the average strength sum(f_i(I))/N and sets the I that gives it (exact inverse of the piecewise-linear curves).
• magnet[i].strength sets I = f_i^-1(x) with the magnet's own submodel; the other magnets follow. Design no longer splits the value by length.
• Array writes give the PS the mean of the currents asked by the series members present in the array, not by all N, so writing one member no longer moves the PS by 1/N. Inconsistent requests log a warning. Design arrays (no aggregator) apply the same rule as live.

Combined-function magnets keep their current behaviour: the strength aggregator now calls MagnetModel.compute_hardware_values_partial(), whose default is the previous logic; LinearSerializedMagnetModel overrides it.

Also:

• Build the serialized submodels once, at configuration. Attaching a family used to rebuild them, so the per-magnet accessors kept stale submodels that never received the magnetic rigidity.
• SerializedMagnetsArray.unit() no longer splits unit strings into characters.
• Remove ReadWriteSerializedStrengths.elements / get_elements(); use SerializedMagnets.get_magnets().

Tests: new design/live tests for idempotence, family values, per-magnet writes, design/live parity and array writes; tests that relied on the sum and length-sharing convention are updated.

Refs #461
Closes #473