Estimate # prompts, # input tokens per probe

[leondz]

Summary

We need to help estimate the inference load a run may bring.

One interface might be to have a .estimate_prompt_count() method on prompts

We first know the number of prompts after the probe has enqueued its attempts, but before these are sent for inference.

It would be worth making this possible to override for probes that are dynamic/adaptive.

Returned values are estimates, and don't need to be precise, but should be useful for e.g. building a run-level tqdm bar. User should expect that estimates are always within an order of magnitude of accuracy, and often within 10% of the real count.