Speed regression for n_jobs=1

[lesteve]

from scikit-learn/scikit-learn#8186 (comment):

Snippet reproducing the issue:

from joblib import Parallel, delayed

def func():
    return 42

inputs = range(1000)

%timeit Parallel(n_jobs=1)(delayed(func)() for _ in inputs)

The speed regression is way worse on Python 2:

Timings:

	Python 2.7	Python 3.5
joblib 0.9.4	32.8 ms	38.6 ms
joblib master	667 ms	76.6 ms

[lesteve]

Profiling, the culprit is the timeout support and in particular this line.