Add support for non-numeric batch sizes

[ruivieira]

At the moment, string-based batch size like auto or auto:N that lm-evaluation-harness accepts is not supported for API models, and it will fallback to the default batch size of 1.

However, any integer batch size that is provided is indeed used for API models. So a few observations:

1. Our current CR does not support non-integer batch sizes, so we don’t support the auto feature either.
2. Our default batch size of 8 I’ve found to be too large in every evaluation