Parse stdout/stderr from forward model steps for Warnings

[berland]

Ert wants to be able log warnings (as in all kinds of warnings) emitted from the forward model steps.

These steps run in separate subprocesses for which Erts root logger have no influence, thus such a feature must rely on text-parsing stdout and/or stderr.

(This is an alternative approach to https://github.com/equinor/scout/issues/1151)

[berland]

There are some open questions to how this should be implemnted:

• On the compute side (in forward_model_runner) or on the scheduler side (scheduler.py/job.py).
• To avoid flooding logging systems with repeated messages, deduplication of warnings must be done on the scheduler side.
• If parsing is done on the scheduler side, pay attention to CPU-cost, and whether one or every runpath should be checked.
• Should a human and machine readable file (xml/json/whatever) be left in the runpath with a list of each warning with metadata (step name, name of parsed file, etc)?
• Should parsing happen constantly, or when a step is finished, or when a realization is finished?

Later, it can be decided if any of this information should be added to a GUI element as in #9381