Implement protobuf format in client_python and add support for native histograms

[m...@alesieur.net]

Hi,

I am aware that there is work being done to approve a plain text format for native histograms (see https://github.com/prometheus/client_python/issues/918 for instance). But it is also my understanding that it's not approved yet, and that it would then need to be implemented in the whole stack before it becomes usable.

As a result, I started working on adding protobuf support to the python client which would open the door to adding native histograms.

I wanted to see what I was fighting against so I already started an implementation, and have a proof of concept for a client that will ship data in protobuf delimited format, similar to what the golang client does.

I'm at the point where it makes sense to gather feedback from this group, to see if this is something that we would be interested to work on and actually merge, in one form or another, into the actual client, or if this is something that will stay a side project for me, and potentially be used at my work while we wait for the plain text implementation.

I can quickly explain what I've done:

- imported the metrics.proto file found in the prometheus repo. I removed references to gogoproto, and added a bunch of `optional` qualifiers to make some edge cases work better (mainly passing existing tests if we're honest)

- the new proto classes are then used extensively in metrics.py and metrics_core.py

- since I was going for minimal changes, I kept the underlying Sample/Timestamp/... classes found in samples.py The obvious benefit for me is that it meant not having to rewrite all the plain text generators, and just having to make sure that I could translate protobuf classes into `Sample`.

- from there, marshalling the new samples into binary strings to send over the wire is fairly trivial

The BIG hurdle in my current implementation is that we currently offer the ability (mainly, hopefully exclusively?, used in tests) to manually override the content of `Metric.samples`. This is of course a problem for me as I'm storing everything as a protobuf `MetricFamily`, and only generate samples on the fly if I need to.

Because I wasn't sure of the implications, I decided (again, as a first POC), to keep a `Metric._samples` list available. You can still manually set it (it's tied to the `.sample` property). Which has the side effect of also clearing all the protobuf metrics currently stored in the object, and if you ask for `Metric.samples` I return both.

There are a couple other side effects. The main one being that `Metric.timestamp_ms` in protobuf is, as the name indicates, in ms (int). Other timestamps (exemplars, created) are stored as ns. It does mean that in my version, the timestamp of a metric is no longer in nanosecond resolution. Which translates to the plain text outputs. There are other small changes in the output for similar reasons (underlying protobuf types) but we can talk about these later.

I feel like that enough text for a first message to this group.

Looking forward to hearing from y'all. I need to clean up my work a bit, but if that's helpful I'm of course happy to push this to a public branch and share it here if requested.

Thanks!