Reset tuned Metropolis parameters in sequential sampling of chains

[michaelosthege]

• add a regression test that detects the issue
• implement the fix for Metropolis chain tuning is differs between single- and multiprocessing #3733
• update release notes

[codecov]

Codecov Report

Merging #3796 into master will increase coverage by 0.31%.
The diff coverage is 100%.

@@            Coverage Diff             @@
##           master    #3796      +/-   ##
==========================================
+ Coverage   90.47%   90.78%   +0.31%     
==========================================
  Files         133      133              
  Lines       20537    20552      +15     
==========================================
+ Hits        18580    18659      +79     
+ Misses       1957     1893      -64

Impacted Files	Coverage Δ
pymc3/tests/test_step.py	100% <100%> (+7%)	⬆️
pymc3/step_methods/metropolis.py	88.83% <100%> (+0.14%)	⬆️
pymc3/distributions/continuous.py	80.85% <0%> (+0.4%)	⬆️
pymc3/tests/test_mixture.py	99.65% <0%> (+0.69%)	⬆️
pymc3/step_methods/hmc/base_hmc.py	95.32% <0%> (+0.93%)	⬆️
pymc3/tests/test_transforms.py	100% <0%> (+1.47%)	⬆️
pymc3/tests/test_distributions.py	98.14% <0%> (+1.72%)	⬆️
pymc3/tests/models.py	85.95% <0%> (+15.7%)	⬆️