Fix and test autograd with hidden units

[kevinchern]

• Fix the autograd computation when hidden units are present. The gradients were computed incorrectly because the linear and quadratic parameters were used to marginalize hidden units without detaching from the computation graph. The fix is to detach them when computing effective fields.
• Add a test to check the gradient is as expected.

Thanks @mhramani for flagging the lack of tests, which lead to discovering this bug!

[mhramani]

LGTM, made a minor comment

[mhramani]

looks like this is not sorted according to h1v1 + h2v2 + h3v3 + J23v2v3 + J12v1v2 + J13v1v3:
I think it should be
q_plus = torch.exp(-torch.tensor(0.2 + 0.2 - 0.3 - 0.6 + 0.2 - 0.3))
q_minus = torch.exp(-torch.tensor(-0.2 + 0.2 - 0.3 - 0.6 - 0.2 + 0.3))
this won't change the results

[kevinchern]

great catch, thanks!

[mhramani]

@kevinchern this is a very important and time-sensitive PR, I already reviewed approved, could you please merge. @thisac Please have a look once you get a chance.

[kevinchern]

Thanks @mhramani