JAX evaluator backend with analytic Jacobians (v0.12.0)

[InfinityMonkeyAtWork]

Adds an experimental JAX backend for 2D fits, opt-in via Project.spec_fun_str =
"fit_model_jax" and the [jax] extra. Implements Phases A–D of
docs/design/jax-planning.md; execution record in docs/design/archive/jax-backend.md.

• Phase A (backend-agnostic): expression programs flattened to packed CSR arrays; plans
  fully array-native, evaluator API unchanged.
• Phases B+C: jitted 2D evaluator covering the full can_lower_2d surface (dynamics,
  expressions, subcycles, profiles, kernel-matrix convolution, Voigt via Weideman wofz,
  <1e-12 vs scipy). Gate composes with can_lower_2d, so the fallback chain is JAX → NumPy
  GIR → MCP by construction.
• Phase D: jax.jacfwd analytic Jacobian wired into lmfit leastsq via Dfun. Voigt 2D fit:
  25.3 s → 7.1 s (~3.6x incl. compiles), identical redchi, nfev 8 → 2.
• Testing: every comparison in test_evaluate_2d.py now asserts three-way parity
  (interpreter/NumPy/JAX) at 1e-12; JAX-specific tests cover gate, Jacobian-vs-FD, wofz,
  e2e fit. Main CI installs .[dev,jax]; min-versions stays jax-free (jax needs numpy≥2.0).
• Known limitations (documented): no runtime value checks on traced params (LinBack
  ordering, kernel positivity), no parallel-worker MCMC, imports enable global x64, 1D fits
  stay on the NumPy path.
• Phase E (optimizer replacement) deferred with rationale; follow-on design notes in
  docs/design/ui.md and docs/design/project-level-fits.md.