Neurodiffeq

Happy 2024 🎆 🥳
- Add support for selecting a particular device when there are multiple GPU devices.
- Fix an bug when loading a solver whose `.nets` have shared instance(s) of `torch.nn.Module`.
- Miscellaneous documentation improvement.
- PR CI is up and running again

1. Allow update best_net without running validation every epoch
2. New 1-D Dirichlet boundary condition for bundles
3. Some API changes for bundle solvers and bundle conditions
4. Allow saving and loading custom solvers inheriting `Sovler1D`, `Solver2D`, or `BundleSolver1D`
5. Suppress error related to removing colorbar caused by newer versions of matplotlib

Hot fix: solve a fatal compatibility issue with torch v1.13

- Bug fix: deduplicate parameters when networks share common layers/params
- Removed deprecated APIs
- Forward compatibility with newer versions of PyTorch

1. Better progress bar
2. Allow user to conveniently set torch and numpy seeds with `neurodiffeq.utils.set_seed`
3. Implement solvers.get_residuals
4. Misc bug fixes and better error messages

- Support Chebyshev points of the first and second kine for `Generator1D`, `Generator2D`, `Generator3D`, `GeneratorND`
- Modify the usage of `log-spaced` and `log-spaced-noisy` for `Generator1D` to make it consistent with other sampling methods
- Support `tqdm` progress bar with `solver.fit()`. Default is writing progress bar to stderr.