Replace AggregateObjective with ScalarizedObjective, drop normalization

[499602D2]

Replaces AggregateObjective with ScalarizedObjective, dropping the flowboost-side outcome-normalization layer entirely. Ax's default transform stack already handles outcome scale (Winsorize → BilogY → StandardizeY at the modelbridge layer + BoTorch's Standardize on the GP); our layer was redundant, and by running post-scalarization it discarded the per-metric information Ax's modelbridge uses.

AxBackend defers to Ax's native ax.core.objective.ScalarizedObjective so each inner metric gets its own surrogate and StandardizeY; other backends would compute the weighted sum at the flowboost layer if they lack an equivalent primitive.

Addresses the concrete AggregateObjective complaints in #8; mixed minimize/maximize inner objectives, different-scale objectives, the absent normalization story, and the batch_process crash reported in the comment. Leaves the broader Pareto-MOO discussion (weighting / thresholds for separate objectives) open.

Migration

# Before
agg = AggregateObjective(
    name="LiftToDrag", minimize=False,
    objectives=[lift, drag], weights=[0.7, 0.3], threshold=None,
)

# After: signed weights encode direction; negative flips drag's contribution
agg = ScalarizedObjective(
    name="LiftToDrag", minimize=False,
    objectives=[lift, drag], weights=[0.7, -0.3],
)

Drop any normalization_step= or attach_post_processing_step(...) from Objective definitions; the backend handles outcome scale. For domain transforms (e.g. log-distributed measurements), use Objective(static_transform=math.log, ...).

Validation

ScalarizedObjective.__init__ rejects at construction:

• Non-finite weights (NaN, ±inf)
• All-zero weight vectors (no signal to fit)
• Duplicate inner objective names (would collide in Ax's metric registry)
• Non-Objective inner terms (nested ScalarizedObjective, Constraint, etc.)
• Inner Objective.threshold (MOO-only; pointless under scalarization)
• Direction/weight-sign mismatch (inner minimize must agree with the direction implied by outer.minimize and the weight's sign)

A single zero weight warns and skips the direction check for that term, so users can temporarily mute a contribution during tuning.

Tests

Unit tests cover every validation branch. A new convergence canary runs BO on scale-mismatched inner metrics and fails if per-metric StandardizeY or the AxScalarizedObjective post-create swap regresses. Existing normalization canaries still pass.