Train/eval imagenet fake data benchmark for a single step.
This fixes #3364 (see comments twoards the end of the issue) - before this change, configurations with a (global) batch size > 512 were failing because of an empty evaluation.
PiperOrigin-RevId: 571455235
