Change tune method to not act inplace plus more sampler stats

[michaelosthege]

By acting inplace, the tune method in metropolis.py modified that scaling of all parallel DEMetropolis chains at the same time.

When DEMetropolis is used with cores = 1, the step method instance is shallow-copied for each chain. The shallow-copying is memory-friendly (important for b/c of the model graph), but the shallow copies also share the reference to the original scaling attribute.

The following example shows why the in-place tune implementation is problematic:

import copy
import numpy

class Chain:
    def __init__(self):
        self.scaling = numpy.array([1,2,3], dtype=float)

chain1 = Chain()
chain2 = copy.copy(chain1)

print(chain1 is chain2)
#> False
print(chain1.scaling is chain2.scaling)
#> True

# In-place operation as previously done by tune method:
chain1.scaling *= 1.1

print(chain2.scaling)
#> [1.1, 2.2, 3.3]

Changes

• added a regression test that checks the tune method to not act inplace
• changed tune method to return a new array instead of returning the modified original
• added accepted and scaling sampler stats for Metropolis and DEMetropolis because they were useful for testing

closes #3731

[ColCarroll]

LGTM too. I restarted the build after clearing the caches on travis.

[codecov]

Codecov Report

Merging #3732 into master will increase coverage by <.01%.
The diff coverage is 100%.

@@            Coverage Diff             @@
##           master    #3732      +/-   ##
==========================================
+ Coverage   89.93%   89.93%   +<.01%     
==========================================
  Files         134      134              
  Lines       20422    20429       +7     
==========================================
+ Hits        18366    18373       +7     
  Misses       2056     2056

Impacted Files	Coverage Δ
pymc3/tests/test_tuning.py	100% <100%> (ø)	⬆️
pymc3/step_methods/metropolis.py	86.79% <100%> (ø)	⬆️