Re-index TorchOptConfig.objective_thresholds from (n_outcomes,) to (n_objectives,) (#5018)

[saitcakmak]

Summary:

Re-index objective_thresholds in TorchOptConfig from (n_outcomes,) with NaN
for non-objective outcomes to (n_objectives,) or None. This is a prerequisite
for supporting ScalarizedObjective as a sub-objective of MultiObjective.

Key changes:

• extract_objective_thresholds returns (n_objectives,) maximization-aligned
  array (sign-flipped for minimize objectives), or None for single-objective.
  Rewritten to use expression-based API (parse_objective_expression /
  extract_metric_weights_from_objective_expr) instead of deprecated
  MultiObjective.objectives / ScalarizedObjective class checks.
• _untransform_objective_thresholds indexes by objective index and un-flips
  the sign when converting back to raw ObjectiveThreshold.bound.
• Replaced get_weighted_mc_objective_and_objective_thresholds with
  get_weighted_mc_objective, which only returns the objective (thresholds
  no longer need transformation so callers use them directly).
• infer_objective_thresholds returns (n_objectives,) maximization-aligned.
• Removed objective_thresholds from SubsetModelData and subset_model
  (thresholds are per-objective, not per-outcome, so subsetting doesn't apply).
• Simplified _objective_threshold_to_outcome_constraints and pruning logic.
• Merged _full_objective_thresholds and _objective_thresholds into single
  _objective_thresholds in the Acquisition class.
• Pass thresholds directly as ref_point to BoTorch input constructors (using
  the new ref_point parameter from D96473523), avoiding the need to convert
  back to outcome space.
• Fixed Pyre type errors in test_acquisition.py for Optional[Tensor]
  return from Acquisition.objective_thresholds.

TRBO isolation (axoptics):

• TRBO uses its own objective_weights and max_reference_point from
  constructor kwargs -- it does NOT use TorchOptConfig from gen().
  Its __init__ expects legacy format: (n_outcomes,) with raw bounds and
  NaN for non-objectives, then handles maximization alignment internally.
• Replaced the extract_objective_thresholds call in axoptics
  _mk_TRBO_generation_strategy with inline legacy logic that produces the
  (n_outcomes,) raw-bounds format TRBO expects, keeping TRBO isolated from
  the new (n_objectives,) format.
• Added comments to trbo.py documenting the legacy input contract and
  future refactoring TODO.

Differential Revision: D96391935

[meta-codesync]

@saitcakmak has exported this pull request. If you are a Meta employee, you can view the originating Diff in D96391935.

[codecov-commenter]

Codecov Report

❌ Patch coverage is 97.33333% with 2 lines in your changes missing coverage. Please review.
✅ Project coverage is 96.28%. Comparing base (ef72a7d) to head (90cf05d).

Files with missing lines	Patch %	Lines
ax/adapter/adapter_utils.py	94.73%	1 Missing ⚠️
...erators/torch/botorch_modular/multi_acquisition.py	0.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #5018      +/-   ##
==========================================
- Coverage   96.29%   96.28%   -0.01%     
==========================================
  Files         611      611              
  Lines       67314    67270      -44     
==========================================
- Hits        64822    64773      -49     
- Misses       2492     2497       +5     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.