fix: token suspicion adjustment

[tazlin]

Once upon a time, before batching and other optimizations, these were the speeds we considered unreasonable but new paradigms, backends and breakthroughs have made these numbers increasingly inaccurate or irrelevant.

While I do think there has to be some sort of longer term solution (such as the one addressing the problem detailed in #463), there have been virtually only false positives, and the few true positives boiled down to innocent misconfigurations.

Further, it appears that certain types of worker-reported failures can artificially inflate token count, which may be its own issue.

For the time being, I am advocating that the number is increased to 100t/s for most models, as recommended by henky, and that we respond to possible abuse of this relaxation with other, more complete and sound, measures.

[tazlin]

and just to clarify, this is to resolve the problem of new workers very consistently being flagged and shadow banned. It happens with shocking regularity and due to the nature of the sanction, goes undetected for long periods of time, reducing good compute and confusing users (who see a model as available while in reality it is not)

Further, I would invite any horde moderator to review the suspicion of any text worker: it will be absurdly high, in the hundreds at least and I have seen several examples of being in the tens-of-thousands. This is almost certainly a direct result of tripping this particular safety and really is a testament to the idea that the values are much too low.