feat: Prometheus support, core entrypoint, packaging improvements

[khvn26]

Overview

• Add Prometheus support
• Cover Gunicorn with request metrics
• Add support for custom application metrics
• Add flagsmith-test-tools pytest plugin and assert_metric fixture
• Add flagsmith start api package entrypoint
• Improve packaging

Prometheus

Instead of using django-prometheus, I decided to introduce the Prometheus client directly, and gradually instrument parts of our stack, starting from the HTTP server. One advantage of this decision is that, unlike django-prometheus users, we are able to label HTTP metrics by request path without performance or cardinality risks. See PrometheusGunicornLoggerMiddleware and PrometheusGunicornLogger code for details.

I've also added a couple of developer tools to help us implement metrics in an organised, robust manner:

• The assert_metric Pytest fixture is installed as part of the flagsmith-test-tools Pytest plugin along with this package.
• The with_labels utility function makes sure you always send a default "unknown" value to Prometheus for missing or null metric values. This greatly simplifies metric usage.
• metrics.py modules contained in your Django application are autodiscovered, which makes it possible to advertise your metrics globally — see flagsmith_build_info for an example.

Existing default metrics and a guideline for adding user metrics are added in the README.

Entrypoint

The library now installs a flagsmith entrypoint, which is essentially a manage.py script that we can modify as needed. I've added a flagsmith start api command that runs the Gunicorn web server, ensuring some important defaults:

• Our custom logger class with JSON and Prometheus support
• Correct support for Prometheus multiprocess mode

The idea is that any flagsmith start subcommand command will accept Gunicorn arguments, as the base endpoints, including health*, version and metrics will always be served by Gunicorn.

The next PR will add an implementation of the flagsmith start task-processor command.

My hope is that we'll be able to eventually replace run-docker.sh with this, which should help us ship binaries for customers that can't support Docker.

Packaging

I've added the following improvements ahead of publishing the package on PyPI:

• Proper PEP 621 compatible packaging configuration
• Classifiers, license information, up-to-date author and maintainer lists
• Loosened dependencies

[matthewelwell]

Added a couple of minor comments, but it looks solid to me.

[rolodato]

LGTM overall, great stuff! I'd like to test how this looks when setting up a Grafana dashboard for these metrics.