Improve logging and remove deadlock

[mfeurer]

• use forkserver context to start processes within the pynisher. This should remove a deadlock due to the use of logging, threading and multiprocessing. As a side-effect, we have to bump the minimal pynisher version to 0.6.2.
• log messages from dask.distributed to a separate logfile
• change a few log message severities or remove the log message alltogether
• randomly try IPs for the logging server inside try/except. Previously, the ID would be determined before starting the server, and therefore the IP could be taken by a different process in the meantime.
• Fix a bug in model deletion (if mem usage too high or max model on disk reached).
• improve output of metadata generation

Failing tests are due to openml/OpenML#1078

[codecov]

Codecov Report

Merging #1012 (869feec) into development (168330e) will decrease coverage by 0.48%.
The diff coverage is 82.50%.

@@               Coverage Diff               @@
##           development    #1012      +/-   ##
===============================================
- Coverage        85.16%   84.67%   -0.49%     
===============================================
  Files              125      125              
  Lines             9820     9826       +6     
===============================================
- Hits              8363     8320      -43     
- Misses            1457     1506      +49     

Impacted Files	Coverage Δ
autosklearn/estimators.py	92.12% <ø> (-0.79%)	⬇️
autosklearn/smbo.py	72.41% <ø> (-10.00%)	⬇️
autosklearn/util/logging_.py	72.63% <55.55%> (-1.06%)	⬇️
autosklearn/automl.py	84.49% <87.50%> (-0.39%)	⬇️
autosklearn/ensemble_builder.py	76.57% <87.50%> (+0.61%)	⬆️
autosklearn/evaluation/__init__.py	80.00% <100.00%> (-2.63%)	⬇️
autosklearn/metalearning/input/aslib_simple.py	95.45% <100.00%> (ø)
autosklearn/metalearning/mismbo.py	100.00% <100.00%> (ø)
autosklearn/util/pipeline.py	76.56% <0.00%> (-6.25%)	⬇️
... and 8 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 168330e...869feec. Read the comment docs.