theory_fn protocol + default CCL backend + fixed pure draw

[cailmdaley]

Introduce the theory-backend protocol, the default CCL backend, and the fixed pure draw in the fork, and migrate the fork's own test suite off the module-level firecrown import.

Scope

datavector.py, param_shifts.py, and a new backend module.

• Theory protocol + default CCL backend. Introduce theory_fn(cosmo_params) -> np.ndarray, aligned to the SACC rows, as the theory entry point; any callable satisfies it. Ship a built-in default CCL backend in a new backend module, replacing the module-level firecrown import (datavector.py:31) and the ParamsMap-through-likelihood path in calculate_concealing_factor. import smokescreen imports no theory backend; pyccl is imported inside the default-backend module at construction.
• Firecrown sidelining. Make the firecrown import lazy — retain the path, remove the module-level import. No firecrown-defaults overlay on the default path.
• Fixed pure draw. Replace the global-seed/order-dependent draw with a local default_rng draw, order-independent, speaking CCL-native parameter names (for example sigma8, Omega_c). No injectable draw_fn abstraction — deferred. The draw fix folds in here: with the injectable abstraction dropped it is a small change coupled to the same datavector rewrite.
• Fork-test migration. Migrate the inherited DESC test suite (which assumes a module-level firecrown import) onto the synthetic and CCL backends.

Acceptance

• import smokescreen succeeds with firecrown absent.
• A synthetic callable satisfying theory_fn blinds a SACC vector end-to-end.
• The default CCL backend blinds a standard cosmic-shear SACC file.
• The draw is order-independent and expressed in CCL-native parameter names.
• The migrated fork test suite passes without firecrown installed.

— Opus on behalf of Cail.