Make op_system callable JAX-native for diffrax/NUTS workflows

[jc-macdonald]

Closes #74

Summary

Adds a backend-aware compile path so compile_rhs (and compile_spec) can return RHS callables that are fully traceable by JAX — no forced NumPy concretization.

Changes

Core (op_system)

• compile_rhs(..., xp=np) / compile_spec(..., xp=None) — new optional xp parameter; defaults to numpy (no behaviour change for existing callers)
• _is_numpy_backend(xp) guard routes array ops through xp.sum(xp.stack(...)) / xp.asarray() for JAX compatibility
• _BackendNamespace / _Indexable internal protocols replace Any in all public callable signatures
• eval_fn skips forced float64 cast in JAX mode (avoids x64 config requirement)

Adapter (flepimop2-op_system)

• New backend: Literal["numpy", "jax"] field on the stepper model
• _get_backend_namespace() lazy-imports jax.numpy via importlib (optional dep, no import-time cost)
• Stepper coerces arrays via NumPy on numpy path; passes through on JAX path

Tests

• test_compile_rhs_with_jax_backend_is_jittable — verifies jax.jit wrapping round-trips
• test_compile_rhs_with_diffrax_solve — ODE solve through diffrax ODETerm
• test_compile_rhs_traces_through_blackjax_nuts — NUTS gradient tracing smoke test
• test_jax_backend_stepper_jittable (adapter) — stepper survives jax.jit
• All new tests use pytest.importorskip and are skipped when JAX/diffrax/blackjax are absent

Packaging

• jax and jax-inference optional dependency groups added to pyproject.toml
• README updated with install instructions and usage examples