Profiles collected by diagnostics should be per component (or process) not per unit

[cmacknz]

We are duplicating the profiles per unit, when the profiles apply at the process level and they should likely be at the component level. For example looking at a log component implemented by Filebeat will show the following where the per unit profiles don't add any value.

tree diagnostics/components/log-default/
├── log-default
│   ├── allocs.txt
│   ├── beat-rendered-config.yml
│   ├── block.txt
│   ├── goroutine.txt
│   ├── heap.txt
│   ├── mutex.txt
│   └── threadcreate.txt
└── logfile-system-b856919c-ade6-4a47-9e0c-28d4822b7ada
    ├── allocs.txt
    ├── beat-rendered-config.yml
    ├── block.txt
    ├── goroutine.txt
    ├── heap.txt
    ├── mutex.txt
    └── threadcreate.txt

[fearful-symmetry]

Running into this while working on #2140, and there's a series of implementation and API issues that kind of lead to this:

1. In implementation, the elastic-agent-client will run any diagnostic callback (including the pprof diagnostics mentioned above) that are registered at the client level, and run them per-unit
2. in the protobuf API between the agent CLI and the agent server, the actual API call is DiagnosticUnits and it operates at the level of individual units; there's no component-level DiagnosticComponents protobuf call or anything.
3. In the protobuf API between the agent and the components, the diagnostic gets compressed down to a single ActionRequest_DIAGNOSTICS type.

So, I feel like the correct way to fix this would be to have a DiagnosticComponents API call, and then leverage the ActionRequest struct in some way so the client can differentiate between unit and component-level diagnostics. At that point we can properly fix the first issue of client-level callback registration.

[cmacknz]

Yeah I think the diagnostics API should have accounted for component diagnostics initially, but since it didn't the next best option is a DiagnosticsComponents API call.

We need a way to support process global diagnostics like CPU and heap profiles and this is the only way that makes sense long term.