Bug#944054: statsmodels: ignored test failures on at least ppc64el

[Rebecca N. Palmer]

The test failures are:

armhf: 86, probably all #924036

i386: 1, TestDFM_Approx.test_smoothed_measurement_disturbance out of 
tolerance: suspect rounding error (x87 extra precision).

arm64, ppc64el, s390x: the same 3, 
base.tests.test_penalized.TestPenalizedPoissonOraclePenalized2*: looks 
like a "find which input variables do something" test is finding 3 when 
there are 4.  This is the kind of task where too-weak real signals can 
legitimately be missed in the noise, but the test input doesn't seem 
that close to this edge: on amd64 it works for mod.pen_weight *= 
approximately 3 to 11 (the test uses 10).  There are also convergence 
warnings.

This may need discussion with upstream.