add transformed_variable node type

greta's four node-type model (data, operation, variable, distribution) doesn't well capture the case when a node is (statistically) notionally a variable, but computationally better treated as an operation. Currently, greta has some cases where a distribution constructor returns an 'operation node following a distribution', e.g. for `wishart()` and `lkj_correlation()` this works, but is a bit hacky. There is some opaque internal code to deal with these situations.

A more rational architecture would be to replace these with a new node type: `transformed_variable`, inheriting from the `operation` node type, but containing all of these different methods. They would essentially be an operation but with a distribution slot, and with methods to reverse the operation (using tfp bijectors), thereby making it easier to set initial values.

This wouldn't be a major change, but would make it make it easier to standardise and understand the internal workings of greta.

The docs and technical details vignette would need to be updated.