Use Clang to build MJX

[DwarKapex]

I got a compilation issue while compiling mujoco-mpc with GNU c++ compiler:

#10 912.6         /opt/mujoco-mpc/mjpc/../mjpc/spline/spline.h:106:9: error: expected unqualified-id before ‘const’
#10 912.6           106 |         const IteratorT<SplineType, NodeType>& other) = default;
#10 912.6               |         ^~~~~
#10 912.6         /opt/mujoco-mpc/mjpc/../mjpc/spline/spline.h:105:37: error: expected ‘)’ before ‘const’
#10 912.6           105 |     IteratorT<SplineType, NodeType>(
#10 912.6               |                                    ~^
#10 912.6               |                                     )
#10 912.6           106 |         const IteratorT<SplineType, NodeType>& other) = default;
#10 912.6               |         ~~~~~
#10 912.6         /opt/mujoco-mpc/mjpc/../mjpc/spline/spline.h:109:68: error: expected ‘)’ before ‘&&’ token
#10 912.6           109 |     IteratorT<SplineType, NodeType>(IteratorT<SplineType, NodeType>&& other) =
#10 912.6               |                                    ~                               ^~
#10 912.6               |                                                                    )

The developers of mujoco-mpc use Clang to compile (https://github.com/google-deepmind/mujoco_mpc/blob/main/.github/workflows/build.yml#L19).
From now on, the default compiler for MJX is Clang, but can be changed via docker arguments.

[yhtang]

Instead of enforcing the compiler toolchain globally, can we set it only for the MPC installation step?

[yhtang]

Why do we need to run pip-finalize twice here? Due to the use of pip-sync in the script, the second run will erase what the first run installed.

[yhtang]

The plan coming out of an offline discussion between @DwarKapex and me is to use multi-stage build to create the MPC wheel using clang and then bring it into the mealkit/final image to avoid interfering with the default compiler toolchain. This is similar to how we build the jaxlib wheel in the JAX core image and also how we build the tensorflow-text/lingvo packages in the T5X/Pax arm64 images.