PyTorch collect model statistics

[albertz]

We should collect some statistics (maybe optionally, configurable) (maybe only every N steps if too costly otherwise).

Of:

• weights
• activations
• gradients of weights
• gradients of activations
• gradients of inputs
• loss (per step; or final on train, dev, devtrain)

Kind of statistics:

• mean, stddev (var)
• min, max, median, other percentiles
• all of the above but for the abs value as well
• all of the above but using momentum (e.g. exp moving average, different decay values)

Also see #1446.