Warn if local replication is dropping a batch

Summary:
As of the refactor of distributed TF Keras (@christiana), it's possible that entire batch(es) is/are discarded if all the batches in an instance's dataset cannot be distributed evenly between its local replicas.

This diff adds a warning to let user know if any batches are being discarded for this reason.

i.e. if you run
```
poprun --nunm-instances 1 --num-replicas 2 python main.py
```
and your (non-repeating) dataset contains 781 batches, each local replica will only use 390 batches and the remaining 1 is discarded.

TF2.5 Only

Reviewers: #tensorflow, #framework_ip_review_-_any_oss_or_third-party_code_use_has_been_approved, markf, christiana

Reviewed By: #tensorflow, #framework_ip_review_-_any_oss_or_third-party_code_use_has_been_approved, christiana

Subscribers: christiana

Maniphest Tasks: T52076

Differential Revision: https://phabricator.sourcevertex.net/D60868

Author: hmellor

tensorflow/python/ipu/keras/extensions/data_adapter.py

@@ -159,6 +159,13 @@ def _infer_steps(self, steps, dataset):
       raise ValueError(
           "Could not infer the size of the data. You must specify the number "
           "of steps to run.")
+    if steps % self._replication_factor:
+      logging.warn(
+          "Dataset of length {} is being evenly distributed between {} "
+          "replicas. The remaining {} batch{} will be dropped.".format(
+              len(dataset), self._replication_factor,
+              steps % self._replication_factor,
+              "es" if steps % self._replication_factor > 1 else ""))
 
     return int(steps // self._replication_factor)