* Deferred initialization of `Bijector`s and `Parameters` is expressed using the `flowtorch.LazyMeta` metaclass
* AffineAutoregressive can operate on inputs with arbitrary `event_shape`s
* A few cosmetic changes like changing `flowtorch.params.*` to `flowtorch.parameters.*`
* Temporarily removed conditional bijectors/transformed distributions and inverting bijectors