Cymr

This version fixes a problem where network recordings did not match the length of the input data.

This version adds a general system for recording model states during simulations. It also makes improvements to recording capabilities of the CMR model and makes internal changes to facilitate future support for intrusions.

* `CMRDistributed` has been renamed to `CMR`, and the simplified implementation of CMR has been removed. The `cymr.cmr.config_loc_cmr` function can be used to quickly set up a simple CMR network that runs using the general CMR framework.
* Model generative simulations are now expected to return items themselves rather than the index of each item in the list. This will allow for models that can simulate arbitrary recalls, including intrusions, rather than being restricted to recalling items from the study phase.
* Models now have general support for recording model state during a simulation, through a `record_subject` method.
* `CMR.record_network` has been removed; use `CMR.record` instead.
* CMR recordings have been streamlined to record one state for each study or recall event. This allows recordings to be easily matched to the input DataFrame.
* API documentation has been expanded to include low-level network operations.

This version changes the API for setting sublayer parameters.

* Setting sublayer parameters is now done by setting a sublayer directly rather than passing a dict for each sublayer.
* All parameter definitions may now be composed in multiple steps using the setting API, rather than having to be set all at once.

This release adds support for sequential evaluation of dependent parameters and fixes some bugs.

* Dependent parameters may now depend on other dependent parameters. Parameters will be evaluated in the order they appear in the Parameters.dependent attribute.
* A problem with using multiple sublayers in CMR-D for generating data has been fixed.
* Sublayer parameters are now printed as part of Parameters objects.

This update adds support for sublayers within a network. Sublayers of context are integrated independently of one another. Sublayers may also have different parameters; for example, different sublayers may have different integration rates. Finally, network weights are specified in a more flexible format, allowing weights to be set using arbitrary expressions.

To implement these new features, changes had to be made to the model API:
* Sublayers must be indicated explicitly using a new Parameters attribute.
* Weights are specified in a different format to allow assigning patterns to any specific region of a connection matrix.
* Previously, special parameters could be used to scale patterns when adding them to a weight matrix. Scaling parameters were normalized before being applied. Scaling parameters are no longer explictly supported, but the same effect can be achieved using dependent parameters (to normalize scaling parameters) and weight expressions (to apply scaling or add an intercept).
* The `Axx` and `Dxx` parameters have been removed from the CMR-D implementation. The same effects can be achieved using weights expressions.

* More features have been added to help work with parameter definitions, including copying, reading from a JSON file, and working with dependent and dynamic parameters.
* Parameter definitions are now passed as a single argument, to make it possible to easily add more options specifying how to evaluate parameters.
* Model evaluation, fitting, and simulation now support dynamic parameters. Dynamic parameters are set for individual study or recall trials and may vary based on both the input data (for example, setting context integration rate to vary by serial position) and other parameters (for example, a slope and intercept to map serial position to the necessary integration rate).