Add mlx-sci (scientific computing toolkit)

[akaiHuang]

Adds mlx-sci under Other Tools.

mlx-sci is a meta-package of nine focused MLX sub-packages providing GPU-accelerated scientific computing on Apple Silicon:

Special functions (mlx_sci.special):

• Spherical Bessel j_l(x) — 579× over scipy at 525 ℓ × 10k pts
• Wigner 3j/6j/9j and Clebsch-Gordan — 2000× batched over sympy
• Gauss hypergeometric ₂F₁ with fused Metal kernel — 7.4× over scipy at N=1M
• Airy Ai/Bi, gamma/lgamma/digamma

Linear algebra (mlx_sci.linalg): matrix expm / logm / sqrtm / Fréchet derivatives.

Signal processing (mlx_sci.signal): class-based STFT/ISTFT layers with mx.compile fusion — 10× over scipy on 30s audio.

Quantum information (mlx_sci.quantum):

• Quantum relative entropy via Stochastic Lanczos Quadrature (650× over NumPy eigh at N=2000; only path that completes past N=2000 on Metal)
• Petz recovery map, fidelity, retrodiction quality
• Quantum channels (thermal attenuator, depolarising, dephasing)
• Fisher information matrix, natural gradient

Install: pip install mlx-sci (PyPI, MIT, Python ≥3.10, MLX ≥0.30).

Each sub-package is independently pip-installable (mlx-bessel, mlx-qre, etc.) for à la carte usage. All headline numbers come with break-even tables and accuracy disclosures in each sub-package's benchmark_results.md.