Replace LDC93S1 sample in prod tests with something else

[reuben]

LDC93S1 prod model outputs vary wildly across versions of the prod model and our inference runtime, because we don't train with anything like TIMIT, so predictions are not confident. We should use some sample from eg. LibriSpeech that will be more stable across changes to prod model and avoid wasting time waiting for test results every time something has to be changed there.

[reuben]

I won't have time to take this for 0.8.