use named io instead of pure vectors

[yunlongxu-numagic]

TODOS before merge:

• [x] add tests
• [x] fix the duplicate of get_group_names and get_cls_group_names
• [x] understand do we need all permutations of apply_and_forward, apply_and_forward_with_arrays',forward,forward_with_arrays`
• [x] currently ModelReturn and StateSpaceModelReturn are used with both arrays and dictionaries value for the entries, need to fix that
• [x] are there ways that we could help the user get better error messages for constructing inputs/outputs? eg, when a user requires an outpuname that doesn't exist, previously self.get_input(inputs, "this_input") would throw some meaningful error, but inputs["this_input"] would only throw key error -> moved to a new issue
• [x] can we/should we move con_outputs outside of model I/O? it is only needed for OCP, not really for models -> move to new issue -> moved to new issues
• [x] can we remove the _dot naming convention for states_dot?
• [ ] pause a bit before the merge, cool down and think about the design while working on something else
• [ ] update README
• [ ] update colab notebook (after release)

[yunlongxu-numagic]

In general it seems to work ok coding wise, by adding a layer of array->dict before the new model io and dict->array after. However what is causing trouble is the jax jit compilation time increases significantly, especially for the hessian:

• with MF52 tires, from 45sec to 160sec
• with Perantoni tires, from 12 sec to 24 sec

execution time is slightly increased, but not so significant.

Why does the jit compilation time increase so much for this thin layer of wrapper?!

[yunlongxu-numagic]

this slowdown seems to be only limited to using docker on MacOS

• using native MacOS + conda env, we get (10sec for 1 tire vehicle, 28sec for 4 tire vehicle, 36sec for named_io)
• this is similar to using docker on linux (GitHub runners), see the benchmark

So:

• native Mac OS much faster than docker on Mac os
• docker on Mac OS also seems to suffer from some unknown issues that significantly exaggerates the jit overhead going from current I/O to named I/O