[Feature request] Make tfp.optimizer.lbfgs_minimize support a list of tensors as starting points.

I feel it is pretty confusing, that the tfp.optimizer.lbfgs_minimize optimizer and the bfgs pendant can only deal with a single real tensor as starting point, while most complex models give you a list of tensors when queried for trainable_variables. I feel it should be consistent with the rest of Tensorflow, since not every model has similar shaped variables, that can be put into a single tensor.

This gets even more confusing, when you take into account, that tfp.math.value_and_gradient() handles lists just fine, be it with a starred parameter function call.

Start with the tfp.optimizer.lbfgs_minimize entry point and compare it with the BFGS counterpart. Review how tfp.math.value_and_gradient handles lists, then define completion as both optimizers accepting lists of differently shaped starting-point tensors consistently.