Cherry-pick of PR #9693 to 6.6 branch. Original message:

So far, the process metricset has been rather strict. If an unexpected error occurred while collecting process information, the whole collection would stop and return an error.

This changes it to keep iterating through processes even when that happens. The unexpected error will be stored in the Process object and sent to Elasticsearch as well as logged as a warning. This only happens the first time the error is encountered for a process, not on subsequent collection cycles (with a typical collection frequency of 1s, that would flood the log and ES).

For error documents, it sets event.kind: error and event.action: process_error.

Fyi, I have renamed ProcessInfo to Process not just because it now contains more than just types.ProcessInfo, but also to bring it in line with Socket in socket.go. Socket already contains an Error field (and that was the inspiration for this change).

Beware: The diff Github shows is misleading in places, it shows replacements/deletions where a few lines have just moved down a bit.

Some additional background on why this change can be found in this comment thread on a PR that introduced some error catching during process collection.

If anybody wants to test what happens with errors, run it as non-root and comment the continue statement in line 375 - it will report errors for processes of other users. At some point, we might want to have a test that simulates an error.