Use Standardize by default for SingleTaskGP (#2630)

Summary:
Pull Request resolved: #2630

X-link: pytorch/botorch#2458

D60080819 recently updated the default `SingleTaskGP` BoTorch priors. One significant change was to remove the use of an outputscale, which may not work well if the outputs aren't standardized. This diff changes the `SingleTaskGP` to use `Standardize` by default if no outcome transforms are specified (this allows users to explicitly pass in `None` if they don't want to use any transforms).

Differential Revision: D60492937

ax/models/torch/botorch_defaults.py

@@ -772,6 +772,7 @@ def _get_model(
             train_Yvar=None if all_nan_Yvar else Yvar,
             covar_module=covar_module,
             input_transform=warp_tf,
+            outcome_transform=None,  # For backwards compatibility
             **kwargs,
         )
     else:

ax/models/torch/botorch_modular/surrogate.py

@@ -358,6 +358,17 @@ def _set_formatted_inputs(
     ) -> None:
         for input_name, input_class, input_options in inputs:
             if input_class is None:
+                # This is a temporary solution until all BoTorch models use
+                # `Standardize` by default, see TODO [T197435440].
+                # After this, we should update `Surrogate` to use `DEFAULT`
+                # (https://fburl.com/code/22f4397e) for both of these args. This will
+                # allow users to explicitly disable the default transforms by passing
+                # in `None`.
+                if (
+                    input_name in ["outcome_transform"]
+                    and input_name in botorch_model_class_args
+                ):
+                    formatted_model_inputs[input_name] = None
                 continue
             if input_name not in botorch_model_class_args:
                 # TODO: We currently only pass in `covar_module` and `likelihood`