Upgrade Pinocchio Python bindings to nanobind

[ManifoldFR]

Description

This pull request adds a new set of Python bindings using the modern nanobind Python bindings framework.

The new set of bindings is a full replacement of the Boost.Python bindings, which will be deprecated in the next major version of Pinocchio.

This PR is the sequel to coal-library/coal#659 on the Coal library, and actually depends on the Coal nanobind bindings as well as nanoeigenpy.

Checklist

• I have run pre-commit run --all-files or pixi run lint
• I have performed a self-review of my own code
• I have commented my code where necessary
• I have made corresponding changes to the doxygen documentation
• I have added tests that prove my fix or feature works
• I have updated the CHANGELOG or added the "no changelog" label if it's a CI-related issue
• I have updated the README credits section

Task checklist

• Math

  • RPY (with submodule)
  • Tridiagonal matrix
  • Linalg (eigenvalue/eigenvector/inverse)
  • Lanczos decomposition
  • Gram-Schmidt

• Spatial

  • Inertia
    • Inertia
    • LogCholeskyParams
    • PseudoInertia
  • SE3
  • Motion
  • Symmetric
  • explog

• Multibody

  • Model
  • Data
  • GeometryModel
  • GeometryData
  • Lie groups
  • Joints
    • Generic JointModel/JointData
    • Variant + base visitor
    • Derived classes (with specific add-ons)
    • Joint Lie groups
  • Pool (class ModelPool...)

• Algorithm

  • ABA
  • ABA derivatives
  • CAT (computeAllTerms())
  • Cholesky
  • COM
  • Centroidal
  • Centroidal derivatives
  • CRBA
  • LCABA (added after Lcaba python bindings Simple-Robotics/pinocchio#32)
  • Constrained dynamics
    • Constraint Cholesky
    • Delassus operators
  • Constrained dynamics derivatives
  • Constraint solvers
    • ADMM
    • PGS
  • Contact dynamics
  • Contact inverse dynamics
  • Contact Jacobian
  • Geometry (updateGeometryPlacements()...)
  • Impulse dynamics
  • Impulse dynamics derivatives
  • Jacobian
  • Delassus
  • Kinematic regressor
  • Kinematics
  • Kinematics derivatives
  • Regressor
  • Model (appendModel(), buildReducedModel(), findCommonAncestor(), transformJointIntoMimic())
  • Energy
  • Frames (updateFramePlacement, framesForwardKinematics...)
  • Frames derivatives
  • RNEA
  • RNEA derivatives
  • Parallel algorithms:
    • Parallel ABA
    • Parallel RNEA
    • bind OpenMPException exception OPTIONAL
    • Util (omp_get_max_threads())

• Collision

  • Broadphase
  • Collision (expose-collision.cpp in old bindings - class ComputeCollision, functions computeCollision[s]()...)
  • "Coal
    • Conversion from coal.Transform3s to pin.SE3 (REVERSE NOT POSSIBLE)
    • Serialization
  • Pool
    • GeometryModel/GeometryData pool
    • Broadphase manager pool

• Parsers

  • URDF
    • Model
    • Geometry
  • SDF
  • SRDF
  • MJCF
  • Graph

• Constraints

  • Generic
  • Variant + base visitor
  • Inheritance visitors (subclasses of BinaryKinematicsConstraint{Model,Data}Base, PointConstraint{Model,Data}Base)
  • Derived classes
  • Cones and sets (Coulomb, Dual Coulomb, Box, zero, full space, non-negative orthant)

• Utilities

  • Type caster for boost::variant (done in pinocchio/bindings/python-nb/utils/boost-variant.hpp)
  • Convert classes (mostly pinocchio/spatial) to numpy arrays using __array__() method, allows

    f = pin.Force()
    f_arr = np.array(f, copy=...)

• Custom call policy to disallow the use of JointMimic. nanobind has call policies.

• Serialization (pretty much everything missing)

• Add Pixi feature and environment

• Stubs (working)

Known limitations

• This PR only adds nanobind bindings for the double scalar type (perhaps float).
• There is no equivalent to the boost::python::convert::implicit struct, which allows to define custom implicit converters between types. There are two avenues for creating implicit converters in nanobind:
  1. {constructor + declare nb::implicitly_convertible/use nb::init_implicit},
  2. creating a type caster.
     This allows defining the coal.Transform3s -> pinocchio::SE3 conversion by registering an extra ctor, but the reverse is not possible.

Other modified

• pinocchio/visualize/meshcat_visualize.py - remove import of DeprecatedWarning, replaced by Python built-in DeprecationWarning

Other removed

• pinocchio/deprecation.py - only used in meshcat_visualizer.py through the DeprecatedWarning
• completely unused header pinocchio/bindings/python/multibody/joint/joint.hpp

Upstream changes

• Add support for Eigen's Tensor module wjakob/nanobind#1320 - Add support for Eigen::Tensor to nanobind 😄
• Nanobind bindings : fix shared_ptr<> for BVH classes coal-library/coal#833 - Fix problems accessing geometry.convex on some Coal BVH types

Notes

• In some of the new Pinocchio 4 stuff we use std::optional whereas we use boost::optional elsewhere (for the bindings mostly):
  • some of the (member) function/ctor overloads in the Pinocchio C++ source currently exist because we couldn't bind boost::optional:
  • we can now in Boost.Python/eigenpy since a while (I added it)
  • we should modernize the source code to use std::optional, and use it everywhere instead of boost::optional
  • nanobind
• Similarly, we should start using std::string_view where appropriate
• The __repr__ method nanobind implements when using nb::bind_vector to bind STL vectors uses the elements' __repr__, so a vector of pinocchio.Model has a VERY long representation.
  • We should distinguish __repr__ from __str__ on some classes.
• I tried to bind the OpenMPException exception in nanobind, inheriting from the mainline Python Exception class using:

  nb::type<OpenMPException>(m, "OpenMPException", PyExc_Exception);

  but the exception class doesn't inherit from one of the STL exception base class and doesn't implement the member function ::what()
• Conversion to NumPy arrays using __array__ was factored into a new visitor (ToNumpyArrayVisitor). Properly supports copies and kwargs now.
• Nanobind is stricter than eigenpy was regarding conversion of nominally 2D (but actually 1D) (n,1) NumPy arrays to Eigen types: it's not allowed, which means code as follows:

  model = ...
  qmax = np.ones((model.nq, 1))
  q0 = pin.randomConfiguration(model, -qmax, qmax)  # TypeError in NumPy → Eigen caster

Resolves #2616