Exemplar fixes

[mdisibio]

What this PR does:
Fixes several issues with exemplars that were found when testing: grafana/grafana#96859

1. quantile_over_time(duration) was returning exemplars in nanos instead of float seconds like the samples. This is fixed as well as for any other function like avg_over_time(duration).
2. Finishes the TODO for the other functions that were capturing placeholder exemplars but without values (always zero):

• rate()/count_over_time() exemplars are attached to the final sample value. See screenshot below
• min/max_over_time() use the value from the span

3. avg_over_time was accidentally dropping exemplars because of struct address issue.
4. Actually randomly samples a span. Maybe it's my test dataset but the current approach of Spans[0] wasn't working well. The first span in my spansets was always the root span, so quantile_over_time(duration, 0.5) was always choosing the span with the longest duration as the exemplar and it was flying outside the graph. Randomly sampling alleviates this problem.

TODOs
This is a lot better but there are still some things we should revisit:

1. span ID? I really want {status=error} | rate() to link directly to the failing spans like search. But for now we aren't reading the span ID column, which would greatly increase query overhead.
2. min/max_over_time() exemplars are kind of funny, they are any random span that was considered. I think the ideal behavior is for the exemplar to be the span that was the minimum or maximum.

Which issue(s) this PR fixes:
Fixes #

Checklist

• Tests updated
• Documentation added
• CHANGELOG.md updated - the order of entries should be [CHANGE], [FEATURE], [ENHANCEMENT], [BUGFIX]

[mapno]

LGTM, nice fixes