Add note about floating point weights in update_weights docs

[arya2]

Motivation

To highlight that update_weights may not return WeightedError::AllWeightsZero when all weights are zero when updating floating-point weights.

[dhardy]

Thanks for the PR. But I don't understand the motivation — is it a warning that you have noticed inaccuracies in the result after using update_weights?

That inaccuracies can occur when using floating point values is one of the basics of Computer Science. Example (which could occur in WeightedIndex::new). How large the error can be is another, much more complex, issue.

As for special treatment for weights which are zero: is your point that total_weight can be trivially reset accurately when all weights are zero? I don't see how this is more useful than periodically replacing the distribution with WeightedIndex::new from calling code.

[arya2]

Is it a warning that you have noticed inaccuracies in the result after using update_weights?

Yep! I should've been clearer about the motivation. It's only meant to help avoid confusion around WeightedError::AllWeightsZero. I made a mistake when reviewing some code and thought it could be helpful to highlight. Perhaps it would be better to update the docs for that error variant instead?

is your point that total_weight can be trivially reset accurately when all weights are zero?

Yeah, it's trivial to do that outside of WeightedIndex if needed so it's probably not very useful.

[dhardy]

Okay, I'll go with this.

Ignoring the CI failure on MIPS (unrelated).