Heatmap binning reducing cardinality for log-scale test points

I recently ran a `LatencyHeatmap` with `input_lengths=[50, 100, 500, 1000]`, for which my actual generated/observed prompt lengths (`num_tokens_input`) ended up being `[76, 136, 656, 1282]` due to the usual tokenizer estimation error / prompt prefixing / etc.

The problem is that because the first two generated values are much closer to each other than the others, the `binning()` function ended up mapping everything to **three** bins: labelled `[226, 528, 1131]` on the plot.

I believe this would also affect similar tests that space their points (pseudo)-logarithmically... Suggested solution options:

1. Since LatencyHeatmap only generates exactly Nbins prompts today, binning them seems needlessly indirect...
- We could update `binning()` to automatically detect if the cardinality of the input `vector` already exactly matches the number of `bins`, and return the data as-is if so with no binning
- It might be nice to also indicate on the plot whether binning was actually done for each axis? E.g. `Input tokens (exact)`, `Output tokens (binned)`
- This would not solve the potential effects of the same issue on output tokens (which would still be binned linearly regardless of user's choice of target values) - but it's not such an obvious blocker there because the cardinality of the data is still high so it's less likely to end up with fewer than target number of bins containing data.
2. Maybe we could provide options for users to toggle whether the bins are linearly or log-spaced? But it would mean more to configure...
3. Since the distribution of the data (both input & output token counts) is likely to be highly non-uniform, might it make more sense to use adaptive binning techniques to represent it? I know the process is more complex & therefore opaque, but the outputs might align more closely with whatever non-uniformly-spaced target values the user set for the test?

Start by tracing LatencyHeatmap into the binning() function and reproduce the issue with input_lengths=[50, 100, 500, 1000] and the observed token counts [76, 136, 656, 1282]. Review the three proposed approaches and define the expected bin cardinality and plot labeling before implementing; done means the chosen behavior is verified for non-uniform test points.

Written by the indexing model from the issue text.