Unify objective_weights to 2D on TorchOptConfig

[saitcakmak]

Summary

Change objective_weights on TorchOptConfig from a 1D (n_outcomes,) tensor to a 2D (n_objectives, n_outcomes) matrix throughout the Torch generation pipeline. The main motivation is to support ScalarizedObjective children inside MultiObjective, where a single objective can be a weighted combination of multiple outcomes — something the old 1D representation could not express.

Key changes across 19 files

New helpers:

• ax/adapter/adapter_utils.py: extract_objective_weight_matrix() — builds the 2D weight matrix from an Objective (one row per sub-objective for MultiObjective, single row otherwise).
• ax/generators/torch/utils.py: extract_objectives() — iterates over rows of the weight matrix to produce per-objective outcome indices and weights.

TorchOptConfig updates (ax/generators/torch_base.py):

• objective_weights is now documented as (n_objectives, n_outcomes).
• is_moo is now a cached_property inferred from objective_weights.shape[0] > 1 (no longer a constructor argument).
• Added outcome_mask cached property: 1D boolean tensor indicating which outcomes participate in at least one objective.

Source updates:

• ax/adapter/torch.py: Produce 2D weights via extract_objective_weight_matrix; use extract_objectives in _untransform_objective_thresholds; drop is_moo kwarg from TorchOptConfig construction.
• ax/adapter/adapter_utils.py: Update get_pareto_frontier_and_configs and hypervolume to use 2D weights.
• ax/generators/torch/utils.py: Update subset_model to operate on 2D weights (column-wise masking/subsetting); update _get_weighted_mo_objective to use extract_objectives; update get_botorch_objective_and_transform to collapse 2D→1D via .sum(dim=0) where a single scalarized weight vector is needed.
• ax/generators/torch/botorch_moo_utils.py: Update get_weighted_mc_objective_and_objective_thresholds, pareto_frontier_evaluator, and infer_objective_thresholds for 2D weights.
• ax/generators/torch/botorch_modular/acquisition.py: Update _update_objective_thresholds for 2D weights.
• ax/generators/torch/botorch_modular/utils.py: Update _objective_threshold_to_outcome_constraints and choose_botorch_acqf_class for 2D weights.
• ax/generators/torch/botorch_modular/generator.py: Update _get_gen_metadata_from_acqf and _build_candidate_generation_options for 2D weights.
• ax/generators/utils.py: Update get_observed and best_in_sample_point to handle the 2D weight matrix.

Test updates:

• 9 test files updated to use 2D objective_weights tensors.

Test plan

• pytest ax/generators/torch/tests/test_utils.py
• pytest ax/generators/torch/tests/test_acquisition.py
• pytest ax/generators/torch/tests/test_generator.py
• pytest ax/generators/torch/tests/test_surrogate.py
• pytest ax/generators/tests/test_botorch_moo_utils.py
• pytest ax/generators/tests/test_torch_utils.py
• pytest ax/generators/tests/test_utils.py
• pytest ax/adapter/tests/test_adapter_utils.py
• pytest ax/adapter/tests/test_torch_adapter.py
• pytest ax/adapter/tests/test_torch_moo_adapter.py

[meta-codesync]

@saitcakmak has imported this pull request. If you are a Meta employee, you can view this in D94287924.