epic: add official Torch and JAX adapters for differentiable EllPHi

[t-uda]

Motivation

Official framework adapters would let users focus on EllPHi's differentiable geometry rather than reimplementing PyTorch autograd.Function and JAX custom_vjp wrappers by hand.

Why This Is Deferred

• It expands maintenance scope through optional dependencies, framework-specific tests, and device/callback semantics.
• The current notebook implementations are a good proof of concept, but not yet a lightweight core-library addition.

Recommended Deliverables

• ellphi.adapters.torch
• ellphi.adapters.jax
• Minimal examples and framework-specific tests

[t-uda]

Triage note (2026-07): #129 (C++ backend for pdist_tangency_grad, milestone ellphi-0.1.3) is substantively upstream of this epic — the Torch autograd.Function / JAX custom_vjp adapters would wrap exactly the pdist_tangency_grad VJP path that #129 accelerates (~174× in the prototype), and the performance of the pure-Python gradient loop was part of why this epic was deferred. Revisit this epic once #129 lands; the public pdist_tangency_grad(coefs) -> (dists, vjp) API is unchanged by #129, so adapter design work is not blocked on its internals.