[FEATURE][BREAKING] Set contact friction for a pair of materials.

[Milotrince]

Description

• scene.add_material(gs.materials.Rigid(...)) returns a handle that entities hold as their material. Passing one
  handle to several add_entity calls shares one surface; geom.set_material gives one entity more than one.
• rubber.set_friction_pair(steel, sliding_friction=0.4) pins the coefficients for that pair of materials in place
  of the maximum over the two geoms. A coefficient left None keeps following the maximum. Re-declaring the same two
  materials retunes them, before or after build.
• Materials stay an authoring concept. Declarations expand onto collision pairs at build, so no material indirection
  reaches a kernel; the override sits on ColliderInfo, keyed by the dense pair index collision_pair_idx already
  provides.
• friction_ratio is per-coefficient: set_friction_ratio(sliding_ratio=..., torsional_ratio=..., rolling_ratio=...) writes only what is passed. The two new arguments are keyword-only. Coefficients and ratios
  read as <mode>_friction and <mode>_ratio throughout.
• array_class.py loses three tensor fields. A geom's and a contact's friction coefficients are each packed into one
  vector, against the one field the override adds.

Breaking:

• entity.material.friction and its torsional and rolling siblings are gone. Read material.options.friction for the
  declared value or geom.friction for what a geom carries.
• get_geoms_friction / set_geoms_friction, the friction-ratio accessors, and RigidSolverState.friction_ratio
  carry the three coefficients on a trailing axis.
• set_geoms_friction_torsional / _rolling are subsumed by the packed setter.
• entity.set_friction_ratio's first argument is sliding_ratio and scales sliding friction alone, where
  friction_ratio scaled all three.

Out of scope: geom_priority, MJCF <pair> import, anisotropic (5-coefficient) friction, sol_params averaging, and
the couplers' own friction rules.

Related Issue

Addresses #2718. The geom_priority field proposed there is not added; a material
pair subsumes it without making users reason about which surface wins.

Motivation and Context

Friction is a property of a pair of surfaces, so the maximum over two geoms is arbitrary: no per-surface value plus a
symmetric rule reproduces a real table. It also hides single-sided changes, since lowering only a robot's friction
against a default-1.0 ground is a no-op. The issue reports a Go2 sweep that reproduced its own baseline that way.

How Has This Been / Can This Be Tested?

pytest tests/rigid/ -v
pytest tests/ipc/ -v -n0          # ipc needs a single worker on one GPU

tests/rigid/test_friction.py::test_friction_resolved_from_material_pair, over n_envs=[0, 2], asserts
deceleration == mu * g for eight boxes on a plane:

• pair coefficients of 0.6 and 0.05 against surfaces that both declare 0.4, so neither a maximum nor a minimum over
  the two reproduces them
• two boxes sharing one material, one taking its material on the geom, one scaled by a sliding ratio
• distinct launch speeds, so the order broadphase reports the boxes in stops matching the order their pairs were
  numbered in

It is separate from test_kinetic_friction because that scene straddles the tipping threshold, where changing any one
coefficient flips an unrelated box between 176 and 0.1 degrees of tilt.

Locally tests/rigid/ is 426 passed and tests/ipc/ 47 passed. test_physics_parity and
test_robot_grasp_fem[external_articulation] fail identically on the merge-base.

Checklist:

• I read the CONTRIBUTING document.
• I followed the Submitting Code Changes section of CONTRIBUTING document.
• I tagged the title correctly (including BUG FIX/FEATURE/MISC/BREAKING)
• I updated the documentation accordingly or no change is needed.
• I tested my changes and added instructions on how to test it for reviewers.
• I have added tests to cover my changes.
• All new and existing tests passed.

Resolves SIM-329

[duburcqa]

I'm not comfortable with modifying array_class. Make sure every single new field is ABSOLUTELY NECESSARY.

[duburcqa]

Friction information are geometry level should be removed if some extract material indirection is added.