Student T tests fail on GPU [bug]

[lsgos]

sample_shape = torch.Size([])

    def rsample(self, sample_shape=torch.Size()):
        shape = self._extended_shape(sample_shape)
        X = torch.empty(shape, dtype=self.df.dtype, device=self.df.device).normal_()
        Z = self._chi2.rsample(sample_shape)
        Y = X * torch.rsqrt(Z / self.df).unsqueeze(-1)
>       return self.loc + self.scale_tril.matmul(Y.unsqueeze(-1)).squeeze(-1)
E       RuntimeError: Expected object of device type cuda but got device type cpu for argument #2 'mat2' in call to _th_mm

pyro/distributions/multivariate_studentt.py:74: RuntimeError

This issue was discussed in #2226 - running make test on the dev branch errors out for me if running on a machine with cuda. I am guessing this hasn't shown up in the CI because it uses a cpu only machine.

I think this bug is pretty simple - it happens because, as we can see in the above snippet, y inherits its device from self.df, and in the fixture, self.df is set to a scalar value. This is not converted into a tensor by the tensors_default_to context manager, and so isn't sent to the gpu.

I fixed this in #2226 by changing the fixture, but @fritzo suggested that it might suggest a missing coercion rather than a change to the fixture, so that change in the PR was reverted and I am opening this issue instead.

[fritzo]

We really do want to test with Python scalars as inputs. This appears to be a bug in MultivariateStudentT which incorrectly coerces df without examining the device of other tenors:

df, = broadcast_all(df)  # this line is buggy