Refactor "beater" package

[axw]

We currently have a big ball of mud in the codebase: the "beater" package and its sub-packages. These comprise:

• our interface to libbeat (including Elastic Agent/Fleet management)
• setup of HTTP & gRPC servers
• definition of HTTP middleware & gRPC interceptors
• route registration (intake, agent config, ...)
• model processor creation
• configuration
• Java agent attachment
• telemetry
• authentication/authorization framework
• self-instrumentation

This is a growing source of complexity and tech debt, manifesting in issues such as #8383. We should look at refactoring "beater" into several distinct modules, with clearly defined responsibilities.

If nothing else, we must define a clear line between libbeat and the rest of apm-server code. This will enable us to better test things like reloading configuration and its effect on monitoring (see issue linked above) without running a complete apm-server. It would also enable us to explore alternative methods of deployment, e.g. splitting APM Server into multiple services to fit a stream processing architecture.