Fix regression #6840

[michaelosthege]

The tune attribute is reset by the sampling iterator.

Closes #6840

Tested locally with

import pymc

def test_issue_6840():
    with pymc.Model():
        pymc.Bernoulli('X', p=0.2)
        idata = pymc.sample(draws=3, chains=2, discard_tuned_samples=False)
    assert idata.warmup_posterior.sizes["draw"] == 1000
    assert idata.posterior.sizes["draw"] == 3
    assert idata.warmup_posterior.sizes["chain"] == 2
    assert idata.posterior.sizes["chain"] == 2

Bugfixes

• Fixes a regression in BinaryGibbsMetropolis and CategoricalGibbsMetropolis

---

📚 Documentation preview 📚: https://pymc--6843.org.readthedocs.build/en/6843/

[codecov]

Codecov Report

Merging #6843 (18e5562) into main (5c600c7) will decrease coverage by 7.66%.
The diff coverage is n/a.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #6843      +/-   ##
==========================================
- Coverage   92.03%   84.37%   -7.66%     
==========================================
  Files          96       96              
  Lines       16378    16376       -2     
==========================================
- Hits        15074    13818    -1256     
- Misses       1304     2558    +1254     

Files Changed	Coverage Δ
pymc/step_methods/metropolis.py	83.95% <ø> (-2.98%)	⬇️

... and 30 files with indirect coverage changes

[twiecki]

Why no test?

[michaelosthege]

I didn't want to increase test runtime with something that tests a non-behavior.

We could start asserting the sizes in other tests that run for all samplers though. Then we'd catch such a bug for all samplers - not just the ones impacted this time.
I'll try to pinpoint such a test case ⏳

[ricardoV94]

Ideally we can extend one of the tests that already runs for all samplers.

In general, the step sampler API seems quite a mess :/
This is just a symptom of it

[oscarbenjamin]

Thanks for the quick fix!