Add nanobind 3.0 split-mode support

[XXXXRT666]

Proposed changes

Update nanobind to 3.0.0 and adopt nanobind split mode for regular python builds, using the nanobind-backend package.

This change:

• builds the regular python frontend against the Python 3.10 Stable ABI, producing an abi3 frontend wheel per platform
• keeps Python 3.13t and 3.14t in linked FREE_THREADED mode and produces specific cp313t and cp314t wheels, since those python releases do not provide a free-threaded Stable ABI so the official backend does not publish wheels for them
• adapts the Python buffer and some low level call paths required by the Python 3.10 Limited API

The public Python API and the existing mlx-cpu, mlx-cuda, and mlx-metal package split remain unchanged.

For context, see wjakob/nanobind discussion #1416.

Checklist

Put an x in the boxes that apply.

• I have read the CONTRIBUTING document
• I have run pre-commit run --all-files to format my code / installed pre-commit prior to committing changes
• I have added tests that prove my fix is effective or that my feature works
• I have updated the necessary documentation (if needed)

[XXXXRT666]

Wheel release matrix

Platform	Architecture	Regular CPython frontend (split mode)	Python 3.13t frontend (linked mode)	Python 3.14t frontend (linked mode)	MLX backend wheels
Linux	x86_64	cp310-abi3	cp313-cp313t	cp314-cp314t	mlx-cpu, mlx-cuda-12, mlx-cuda-13
Linux	aarch64	cp310-abi3	cp313-cp313t	cp314-cp314t	mlx-cpu, mlx-cuda-12, mlx-cuda-13
Windows	x86_64	cp310-abi3	cp313-cp313t	cp314-cp314t	mlx-cpu, mlx-cuda-12, mlx-cuda-13
Windows	arm64	cp311-abi3	cp313-cp313t	cp314-cp314t	mlx-cpu
macOS	arm64	cp310-abi3 for each OS	cp313-cp313t for each OS	cp314-cp314t for each OS	mlx-metal for each OS

The abi3 wheel covers all supported GIL python versions from python 3.10, while 3.13t and 3.14t require specific wheels. The previous release matrix built 48 frontend wheels: 27 for Linux and Windows, and 21 for macOS across 3 OS versions. The new matrix builds 21 frontend wheels: 12 for Linux and Windows, and 9 for macOS. This reduces the frontend wheel count from 48 to 21.

The 13 existing mlx-cpu, mlx-cuda, and mlx-metal backend wheels remain unchanged. Including those backend wheels, the complete MLX release matrix is reduced from 61 wheels to 34.

[zcbenz]

While the split-mode is very helpful for many projects, it does not really make much difference in MLX because we separate frontend/backend packages and the cost of publishing frontend package for every python version is very small for us.

So I think we can keep this open in case we need it some day, but at the moment I don't want to risk breaking things for almost no benefits.