[avgvi]

Under per_stream and device_per_stream batching, a stream that cannot fill a single batch raised ValueError, although the message itself said "Training will occur without any samples from this stream". The whole epoch failed because of one small stream.

Since #57, a stream with fewer samples than num_canonical_nodes no longer fails in the shuffle, but it then reached this check. With the default num_canonical_nodes (the number of nodes), such a stream puts one sample on each node. That fills a device batch only when the batch size is 1, so on more than one rank or with larger batches the epoch still failed. For example, a 1-sample stream next to two large ones, on 2 nodes × 2 ranks with batch size 4, or 2 nodes × 8 ranks with batch size 16.

This change:

• per_stream: a stream that cannot fill one complete global batch is left out of the epoch, with a warning.
• device_per_stream: a stream that cannot fill one device batch on a node is left out on that node, with a warning. Nodes with different batch counts are already padded to the same count, as before.
• Neither method: if no stream can fill a batch (on a node, for device_per_stream), the epoch fails with an error that says so.

Tests:

• test_stream_too_small_for_a_batch_is_skipped runs for both methods on 2 nodes × 2 ranks with batch size 4, with streams of 400, 300 and 1 samples. It checks for the warning and that the 1-sample stream does not appear. It also checks that the other two streams do appear, and that every device batch still comes from a single stream.
• test_no_stream_fills_a_batch checks the new error.
• All four cases fail on main. The existing batching tests pass.