Clean up oxql histogram responses

[jmcarp]

Related to #9324, but distinct.

For histogram metrics, the OxQL endpoint returns a quantile for the p50, p90, and p99 of the histogram, in addition to bucket boundaries and counts. This type is a bit opaque to the end-user, who might not be familiar with the P² algorithm, and includes docs about implementation details that I'm guessing should be internal. Here's an example api response:

...
"p50": {
  "desired_marker_positions": [
    67113772.0,
    4.25,
    7.5,
    10.75,
    14.0
  ],
  "marker_heights": [
    65172855.0,
    98236867.20436507,
    114319734.61547618,
    139226215.7027778,
    163392545.0
  ],
  "marker_positions": [
    1,
    4,
    7,
    11,
    14
  ],
  "p": 0.5
}
...

I think the relevant part of this is that we've estimated the p50 of this series to be 114319734.61547618, the third of the five markers. Although maybe it's also useful to know the p/5 and (1+p)/2 quantile estimates as well? The first and final markers, which represent the min and max of the series, seem to be redundant, since we extract them into their own fields in the parent Histogram struct.

I would like to do two things here:

• Simplify/clarify the response type for histograms, so that we only include useful information, and so that users can tell what it means. This might mean just showing the quantile estimates, or hoisting them into Histogram for clarity if there's value in including the full set of markers in the response.
• Updating the docs to explain histograms better. We're providing users with two different ways to estimate quantiles: estimates of a few specified quantiles via the P² algorithm, as well as buckets and counts that allow users to calculate arbitrary quantiles, e.g. using linear interpolation. I think we can do more to explain this to users, and when they might want to use different parts of the histogram response.

cc, as usual, @bnaecker. I'm happy to make any of these changes if they make sense, or not if they don't.

[bnaecker]

This sounds good. I don't think the user needs to know anything other than the actual estimated percentiles. Removing all the details about markers positions, heights, etc is good.

On a much longer-term note -- while I think the P2 algorithm is very cool, we will probably also want to add methods into OxQL itself for estimating other percentiles, something like a function percentile(histogram_timeseries, 0.75) or whatever. That should also apply to scalar histograms, and effectively be a reduction over some temporal bin-size. This is a pretty big project, but just wanted to throw it out there.