Allow to deep merge custom config with default one for model training and inference

[daniel-wer]

Detailed Description

During AI inference and training, there is an advanced mode which allows to specify a custom workflow config. This is useful if one wants to run a totally different workflow than the default one. However, quite often only one or two small changes are needed and having to specify the full config to accomplish that is cumbersome. Therefore, as an alternative it should be possible to specify just a partial config that will be deep-merged with the default one, allowing to add or overwrite specific keys.

For example:

segment:
  config:
    segmenter:
      absolute_affinity_threshold: -2

to segment using an affinity threshold.