Centralise OTEL metrics from OpenAI responses

[dylanratcliffe]

Since we're calling OpenAI in multiple locations, it would be good if we could centralise the helper functions: https://github.com/overmindtech/api-server/blob/main/server/risks/shared.go#L81-L127

This would mean it would be standardised across all usage

Extra Credit: Perplexity

As per this article: https://medium.com/@furqanshaikh/measuring-llm-confusion-9529a4b5e907

It's possible to calculate the perplexity of a given response. This is basically a metric for how confused a model is. This would be an interesting metric when comparing one prompt against another. It also makes me wonder if there are existing OTEL integrations that would calculate this and potentially other metrics for us easily

[dylanratcliffe]

Fixed in #320