Long evaluation in _pre_test hangs search

[josepablocam]

Thanks for the great tool. When using it, I ran into a bit of
an edge case. For some operators, the
reasonable _pre_test decorator (tpot/decorators.py)
can actually take an arbitrarily long amount of time, effectively
hanging the search (since it never moves past generating candidates).
The small sampled down pretest_X might be reasonable, but
for some operators complexity scales with columns (and it may be
hard to determine the correct number to sample, if any).

You can reproduce the issue by using the following example:

> cat example.py

import tpot
import numpy as np
import pmlb

X, y = pmlb.fetch_data("Hill_Valley_without_noise", return_X_y=True)

clf = tpot.TPOTClassifier(
  max_time_mins=1,
  max_eval_time_mins=1,
  cv=2,
  config_dict={
      'sklearn.decomposition.SparsePCA':{},
      'sklearn.linear_model.LogisticRegression':{},
  },
  verbosity=3,
)

clf.fit(X, y)

This will timeout, rather than finish appropriately

timeout 5m python example.py # will hang (and get killed 5 minutes after)
echo $? # 124 -> timeout

Possible fix

I have a simple patch that adds a timeout to the
.fit call taking place at line 88 in tpot/decorators.py. I can PR it if this seems useful.

[weixuanfu]

Hmm, nice suggestion! Please submit a PR for this issue. Thank you!