GRBM weights are initialized arbitrarily and aren't robust

[kevinchern]

@VolodyaCO brought this up at https://github.com/dwavesystems/dwave-pytorch-plugin/pull/47/files#r2577897489 .
Empirically, the current weight initialization scheme tends to result in optimizer getting trapped in poor local optima.
We should run a simple experiment to pick a better initialization scheme.
For what it's worth, I've found initializing weights to 0 to be quite robust.

[VolodyaCO]

For the problem of GRBM learning a data distribution, zero initialisation is good.
When used in a DVAE, if GRBM training is disallowed at the beginning of training, the encoder needs to learn the GRBM distribution; if zero-initialised, this will make the encoder learn a uniform distribution, which is not great (for the purpose of having the GRBM do something).

I would suggest zero biases and small couplings for the initialisation of GRBM.