Casadi

Changes

Major
- remove ability to change the order of primal-dual variables: now it is fixed to `x`, `lam_g`, `lam_h`, `lam_lbx`, `lam_ubx` (that is, in order, the primal variables, equality constraint multipliers, inequality constraint multipliers, and lower- and upper-bound multipiers)
- introduced new example on sensitivity computation
- lower- and upper-bounds on ``x`, i.e., `csnlp.Nlp.lbx` and `csnlp.Nlp.ubx`, are now masked arrays: this allows to filter out redundant constraints (constraints which will never be active, with either `+inf` or `-inf`)
- made all dual variables be vectors instead of having the same shape as the original constraint

Minor
- better type hinting, docstrings and testing
- fixed error due to `cs.solve` being called with non-square matrix in `wrapper.NlpSensitivity`
- fixed error on valid bound checking when creating new nlp's variables
- added method `csnlp.core.data.find_index_in_vector`

Changes

Major
- `multistart.StackedMultistartNlp`
* fixed bug where slack variables for soft constraints were created twice
* fixed bug where parameters and variables with the same name would overwrite each other
* fixed bug where the values of a multistart solution were still symbolic (instead of numerical)
- improved performance of multistart nlps: solutions are now processed when all are fetched; otherwise, only the best is processed
- added the possibility to pass to `Nlp.solve` and `MultistartNlp.solve_multi` incomplete initial conditions, i.e., where not all the variables of the problem are supplied with a starting condition. By default, missing variables are initialized to zero.
- added utilities for easy retrival of solver's docs under `util.docs.py` (inspired by [mpctools](https://bitbucket.org/rawlings-group/mpc-tools-casadi/src/master/mpctools/util.py))

Minor
- better type hinting and docstrings
- improvements to `nlp.debug`
- changes to default behaviour of `util.plot.spy`
- improved the use of `casadi.vertcat`, `vcat` and `veccat`

Changes

Major
- added `np_random` to `csnlp.util.random`
- implemented random and structured generators of initial start points for NLP multistarting, e.g., `csnpl.multistart.RandomStartPoints` and `StructuredStartPoints`

Minor
- updated readme
- removed usage of `dataclass` in favor of `NamedTupe` in `csnlp.core.debug.py`
- corrected some comments in examples

Changes

Major
- refactored existing class and added a new one for multistart NLPs, i.e., `csnlp.StackedMultistartNlp` and `csnlp.ParallelMultistartNlp`. The former solves the original NLP by stacking copies of it into a single, larger optimization problem; the latter uses `joblib` to parallelize the solver
- better symbolic computations in `csnlp.wrappers.NlpSensitivity`

Minor
- improved representation and ordering of `csnlp.Solution` class
- improved tests

Changes

- in MPC, added `input_spacing` (allows to reduce free inputs by keeping them constant)
- added several compression methods to `util.io.save` and `load`
- added `util.math.repeat`

Changes

Major
- added new properties to `Mpc` and `Solution`
- improved `util.math.prod`
- simplified `util.io.SupportsDeepcopyAndPickle`

Minror
- remove functions with dependency from `scipy`
- better type hinting
- added slots to most classes