Jax-fdm

Added

Support for differentiable CPU sparse solver
- Added support for differentiable CPU and GPU sparse solvers to compute the XYZ coordinates of the free nodes with FDM. The solver is custom made and registered via a `jax.custom_vjp`. The forward and backward passes of the sparse solver were made possible thanks to the contributions of denizokt. This solver lives in `equilibrium.sparse.py`.
- Added `spsolve_gpu` and `spsolve_cpu`. The sparse solver backend is different (the former uses CUDA and the latter scipy) and it gets selected automatically based on what the value of `jax.default_device` is at runtime with function `register_sparse_solver`.
- Implemented `EquilibriumStructureSparse`.
- Division of responsabilities: created a `EquilibriumModelSparse` that solves for static equilibrium using an `EquilibriumStructureSparse` and the sparse solver. `EquilibriumModel` implements the dense linear solver.

Changed

- `LossPlotter` instantiates a dense `EquilibriumModel`.

Removed

Added

Changed

Removed

- Removed `DTYPE` from `combine_parameters` signature in `jax_fdm.parameters` due to clashes with JAX versions on windows.

Added

Changed

- Added `ParameterManager.indices_opt_sort` to fix bug that mistmatch optimization values for unconsecutive parameter indices. This bug applied to individual parameters and to parameter groups.

Removed

Added

- Implemented parameter groups!
- Added `EdgesForceDensityParameter`.
- Added `NodesLoadXParameter`, `NodesLoadYParameter`, `NodesLoadZParameter`.
- Added `NodesSupportXParameter`, `NodesSupportYParameter`, `NodesSupportZParameter`.
- Added `EdgesForceEqualGoal` to goals.
- Implemented `area_polygon` in `jax_fdm.geometry`.
- Added `FDNetwork.number_of_supports()` to count number of supported nodes in a network.
- Added `network_validate` to check the validity of a network before equilibrium calculations.

Changed

- Sped up `EquilibriumModel.nodes_free_positions()` computation after replacing `jnp.diag(q)` with `vmap(jnp.dot)(q, *)`.
- Vectorized error computations in `LossPlotter.plot()` to expedite method.
- `OptimizationRecorder.record()` now stores history in a dictionary, not in a list.
- Fixed bug in `FDNetworkViewerArtist` that ocurred while plotting reaction forces on unconnected nodes.
- Turned `TrustRegionConstrained` into a first order optimizer.

Removed

- Removed `LossPlotter.network` attribute.
- Replaced `NodeAnchorXParameter` with `NodeSupportXParameter`.
- Replaced `NodeAnchorYParameter` with `NodeSupportYParameter`.
- Replaced `NodeAnchorZParameter` with `NodeSupportZParameter`.

Added

- Added `Goal.is_collectible` flag to determine if a goal should form part of an optimization `Collection`.
- Implemented `EdgesLengthEqualGoal`.
- Added `AbsoluteError` term.
- Added `MeanAbsoluteError` terms.

Changed

- Fixed omission of `gstate.weight` when computing `PredictionError`.

Removed

Added

- Implemented `LBFGSBS` to ensure compatibility of JAX gradients with scipy.

Changed

Removed