Project-level shared fits on the fused JAX backend (v0.13.0)

[InfinityMonkeyAtWork]

Summary

Adds a compiled fast path for project-level shared-parameter fits: when
Project.spec_fun_str == "fit_model_jax" and every file's model passes the
can_lower_2d + can_lower_jax_2d gates, Project.fit_2d() runs one fused,
jitted JAX program over all files (per-file plan trace, static fit-window
slices, flattened + concatenated residuals) with an analytic joint Jacobian
via jacfwd wired into lmfit's Dfun.

• New spectra.pack_project_theta (combined-theta gather arrays) and
  eval_jax.make_project_evaluator_2d_jax / make_project_jacobian_2d_jax.
• Whole-project fallback to fit_project_mcp if any file misses a gate or
  jax isn't installed — no mixed backends; the backend in use is reported in
  the fit output.
• Heterogeneous energy/time grids across files are supported (per-file
  static window slices, concatenated residual vector).
• Measured speedups vs the interpreter: 2.6x (2 files) → 7.6x (4) → 17x (8);
  XLA compile stays sub-linear (< ~0.4 s at 8 files).

Tests

• TestProjectFitJax: interpreter/JAX end-to-end parity, truth recovery,
  heterogeneous grids, fallback gating (non-lowerable file, jax missing).
• Full suite: 978 passed; sphinx -W docs build clean.

Docs / release

• Design note docs/design/project-level-fits.md updated with implementation
  status and measured results; supported_models.md and
  roundtrip_test_matrix.md updated for the new backend axis.
• CHANGELOG.md 0.13.0 entry; version bumped to 0.13.0. Tag v0.13.0 after
  merge to trigger the PyPI release workflow.