Metrics: Support configurable prometheus histogram bucket sizes for request duration metric

[samlaf]

See #94 (comment)

Metrics for request durations currently have a hardcoded histogram bucket size, which is currently tailored towards eigenda latencies).
Would be nice to let users set them as needed, such that different configurations of the proxy server could use different buckets sizes (memstore, s3, and eigenda backends have very different latencies).