[QoL] Add generated text summary to top of output file

[benchislett]

Overview

Include the generated_text section from each request in the top of the output-file. This makes it much easier to check the response quality at a glance, and correspond them by line-number to the i-th prompt.

Other

• Allow small output token lengths as long as they are < 1. This was formerly a sanity check for ShareGPT data but isn't really needed anymore. Now we can run with --output-token-distribution same 1 to test prefill-only workloads.