Probflow

* stnkl fixed a bug in `probflow.utils.ops.rand_rademacher` (it was previously generating all 0s 😬)
* Fixed some bugs in tests relating to the newest version of PyTorch.

* Allow EarlyStopping to just take a MonitorMetric or MonitorELBO directly
* Remove `expected_calibration_error`, and replace with a more general `calibration_metric`, which can compute any of several different calibration metrics (like mean squared calibration error, mean absolute calibration error, miscalibration area, etc).
* Add `sharpness` and `dispersion_metric` methods to `ContinuousModel` (secondary uncertainty estimate metrics)
* Write the callbacks user guide section
* Some minor callbacks improvements
* Add `CenteredParameter`, which creates a vector of parameters which is constrained to have a mean of 0 (or a matrix whose rows and/or columns are constrained to have means of 0)

* Add `batch_norm` keyword argument to `DenseNetwork`, which can be either `True` (use batch normalization between layers) or `False` (do not use batch normalization between layers, the default)
* Add `batch_norm_loc` keyword argument to `DenseNetwork`, which can be either `'after'` (apply batch normalization after each layer's activation function) or `'before'` (apply batch normalization before each layer's activation function)

* Implement `probflow.distributions.Mixture` for PyTorch.

* Add the `n_mc` kwarg to `probflow.Model.fit`, which sets the number of monte carlo samples which are taken per batch. This performs parameter updates using average of the gradients across multiple MC samples per batch. It's slower with more samples, but leads to more stable fitting.

* Add support for flipout with the PyTorch backend.
* Add `randn`, `rand_rademacher`, and `shape` backend-independent ops
* Update deprecated `tfp.python.math.random_rademacher` in favor of `tfp.random.rademacher` when possible