See: https://ci.appveyor.com/project/tdaff/ase/build/job/u6yxt7e0ifkq44k2
The cumulation of errors in the different dos caluculation methods using 32-bit python (probably in numpy) goes slightly over the current test precision (it's about 1.06e-14):
assert abs(dos1 - dos1w).max() < 1e-14
Increase the tolerence slightly to allow the tests to pass.