Fix backwards compatibility bugs with UVPSpec data files

[philbull]

Fix two backwards compatibility bugs related to loading old HDF5 files into UVPSpec objects.

The first is related to HDF5 in Py3, which returns bytes instead of strings. A string comparison was failing as a result. All string attributes in UVPSpec files are now converted from bytes to strings.

The second is due to the recent change from pol_array to polpair_array in the UVPSpec data model. The old pol_array attribute is now detected and automatically converted to a polpair_array when a file is loaded. A UserWarning is shown when this happens.

[philbull]

This should fix issue #194.

[coveralls]

Coverage decreased (-0.09%) to 96.516% when pulling ae0307e on uvpspec_backwards_compat into 0bee9d2 on master.

[nkern]

hera_qm and hera_cal, although they don't strictly require it, have some advanced functionality that relies on scikit-learn, so it'd be nice to keep this if possible. was something breaking with the installation?

[philbull]

Thanks Nick. The reason from removing scikit-learn from this line is that it was pulling in an old version of numpy, which is now incompatible with pyuvdata (thus causing the tests to fail). It is installed via pip later on in the Travis script however (pip allows it to install with a more recent version of numpy).

[nkern]

Looks good to me, but I'd like to keep scikit-learn installation if its not breaking anything

[plaplant]

Thanks for the quick turnaround @philbull!