Revert batch size gradient scaling fix

Remove the losses.sum() change since it was split into standalone PR #120.
This PR should focus only on dtype utilities.

Co-authored-by: Lucia Quirke <luciaquirke@users.noreply.github.com>

Author: github-actions[bot]

bergson/collector/collector.py

@@ -565,7 +565,7 @@ def fwd_bwd(model, batch):
             if "advantage" in batch:
                 losses *= torch.tensor(batch["advantage"], device=losses.device)
 
-        losses.sum().backward()
+        losses.mean().backward()
         model.zero_grad()
 
         return losses