Use expm1 in Bilog outcome transform (pytorch#2541)

Summary:
Pull Request resolved: pytorch#2541

Similar to pytorch#2540, replaces `exp(x) - 1` with the numerically more stable `expm1`.

Reviewed By: Balandat

Differential Revision: D62813775

fbshipit-source-id: 3e6b7730e8db0d089c123a3477885593f4c7c2ad

botorch/models/transforms/outcome.py

@@ -790,7 +790,7 @@ def untransform(
             - The un-transformed outcome observations.
             - The un-transformed observation noise (if applicable).
         """
-        Y_utf = Y.sign() * (Y.abs().exp() - 1.0)
+        Y_utf = Y.sign() * Y.abs().expm1()
         outputs = normalize_indices(self._outputs, d=Y.size(-1))
         if outputs is not None:
             Y_utf = torch.stack(
@@ -822,5 +822,5 @@ def untransform_posterior(self, posterior: Posterior) -> TransformedPosterior:
             )
         return TransformedPosterior(
             posterior=posterior,
-            sample_transform=lambda x: x.sign() * (x.abs().exp() - 1.0),
+            sample_transform=lambda x: x.sign() * x.abs().expm1(),
         )