Cannot set posterior using RVPair

By specifying the second argument to RVPair, it should be possible to set posterior of a distribution used in a module.

For example (based on a test from borch / a conversation in Zulip):

            module.weight = RVPair(
                dist.Normal(0, 5),
                dist.Normal(
                    Parameter(torch.tensor(2.0)),
                    borch.Transform(
                        torch.nn.functional.softplus, Parameter(torch.tensor(0.1))
                    ),
                ),
            )

Should set the value of the posterior for weight to 2.0. However, the value is unchanged. The following test reproduces the issue:

def test_setting_posterior():
    model = borch.nn.Linear(4, 4)
    ones = torch.ones(4, 4)
    model.weight = borch.RVPair(
        borch.distributions.Normal(ones, 5),
        borch.distributions.Normal(
            torch.nn.Parameter(ones),
            borch.Transform(
                torch.nn.functional.softplus,
                torch.nn.Parameter(torch.tensor(0.1)),
            ),
        ),
    )

    assert model.prior.weight.loc.sum() == 16  # Works
    assert model.posterior.weight.loc.sum() == 16  # Doesn't work

A quick fix is to manually set the posterior after using RVPair:

def test_setting_posterior():
    model = borch.nn.Linear(4, 4)
    ones = torch.ones(4, 4)
    model.weight = borch.RVPair(
        borch.distributions.Normal(ones, 5),
        borch.distributions.Normal(
            torch.nn.Parameter(ones),
            borch.Transform(
                torch.nn.functional.softplus,
                torch.nn.Parameter(torch.tensor(0.1)),
            ),
        ),
    )
    model.posterior.weight = borch.distributions.Normal(
        torch.nn.Parameter(ones),
        borch.Transform(
            torch.nn.functional.softplus,
            torch.nn.Parameter(torch.tensor(0.1)),
        ),
    )

    assert model.prior.weight.loc.sum() == 16  # Works
    assert model.posterior.weight.loc.sum() == 16  # Works

Code should be added in the relevant place(s) to get the expected behaviour without manually setting the posterior afterwards.