feat(solver): add dynamical bias hook and filtered state to CaSuDaSolver

[gcattan]

Step 1 of the dynamical / RTRBM direction. Backward-compatible, opt-in.

• bias_func(run, m, m_filt): per-timestep hook that replaces the static h, enabling recurrence (RTRBM) and adaptive biasing. Defaults to None, in which case behavior is bit-for-bit identical to before.
• m_filt: leaky-integrator companion state (graded [-1, 1]), controlled by a new solver tau param. It carries step (run-1) when bias_func is called, giving the hook temporal memory; never affects the hard +/-1 sampler unless a bias_func consumes it.
• return_filtered: opt-in to append the m_filt trajectory to the return; the default return contract is unchanged for both n_shots==1 and n_shots>1.
• Energy E now uses the effective (biased) h so it stays consistent with all_I when a bias_func is active.

Adds tests/test_dynamical.py covering backward compatibility, filtered-state bounds, the hook feeding the dynamics, energy-reflects-bias, multi-shot shapes and per-shot broadcasting, and tau=0/tau=1 filter invariants.