Moving Hermite basis 1/5: carry u_s, alpha_s in the solver state

[rogeriojorge]

Adds an optional basis state stack([u_s, alpha_s]) (shape (2, 3*Ns)) to the solver state. It is the first of five PRs for a moving and rescaled asymmetric Hermite basis.

• ode_system accepts (Ck, Fk, basis). The state's u_s, alpha_s replace the configured ones in the kinetic operator and in the current.
• simulation(frame="fixed") carries the basis and returns basis_u and basis_alpha per saved time. diagnostics uses the per-time basis for the kinetic energy.
• The default frame=None is unchanged.
• The basis is stored as complex with zero imaginary part, because Diffrax needs one dtype across the state.

Tested SHA 11e6b6e. Command: python -m pytest -q -n 2 tests. Result: 154 passed (baseline 149).

New tests (tests/test_moving_basis.py):

• RHS with the basis in the state and zero rates is bitwise equal to the fixed-basis RHS. Setup: 3 species, unequal orders (Nn, Nm, Np) = (4, 3, 2), a 3D grid, all six field components, nu != 0.
• The state basis overrides the configured one, bitwise.
• J and kinetic energy with a per-time basis agree with Gauss-Hermite quadrature to 1e-13.
• A constant-step trajectory is bitwise equal. An adaptive run agrees to 1e-9 relative: the RMS error norm also counts the basis components, so step sequences differ.

Stack: base adaptive/base = integration/dark-baseline 9d0982d (main ab87385 + #48, #46, #13, #12, #9, #55, #50, #44, #18, #51). This lands on main after those foundation PRs and the earlier PRs of this stack. Local results only (macOS, Python 3.12, JAX 0.11.2, x64); GitHub CI on this branch has not been checked yet.
Formulation and literature checks: Pagliantini et al. arXiv 2208.14373; Bessemoulin-Chatard and Filbet; Ma, Sun and Tang.