Fix the appending of unit suffix for histogram metrics in Prometheus exporter.

[tobz]

In #535, support was added to metrics-exporter-prometheus for suffixing metrics based on their configured unit. This was applied uniformly to counters, gauges, and histograms, and relied on a change to one of the core functions used to write out a metric line in the exposition format.

#552 was raised to highlight a bug with how that change was gated, which made me realize that we're likely suffixing histograms wrong when the feature is enabled. For counters and gauges, the suffixing we do is fine: my_counter turns into something like my_counter_seconds, and so on. For histograms, however, lines like my_histogram_sum turn into my_histogram_sum_seconds.

As I understand it, at least, this is wrong, since the intent is that Prometheus would expect the base metric name to be turned into <base>_bucket, <base>_sum, and <base>_count... whereas we would be doing something like <base>_bucket_seconds, <base>_sum_seconds, and <base>_count_seconds. While the suffixing is uniform, Prometheus wants (again, as I understand it) a histogram metric to have the _bucket/_sum/_count suffixes so that it knows how to transforms queries against <base> into the right parts when doing histogram-y things.

Looking at the code, this should be as simple as reordering how the unit suffix and the generic suffix (the parameter we use to add _bucket/_sum/_count in the first place) are applied.