Fix ESS-target comparison in adaptive persistent sampling solver

[thomasckng]

The root-finding function in adaptive_persistent_sampling.calculate_lambda compared ESS and target in mismatched spaces. compute_persistent_ess returns ESS in linear space, but target_val is computed as jnp.log(n_particles * target_ess) (log space). The original code subtracted the log-space target directly from the linear-space ESS:

return ess_val - target_val

This meant the solver was finding a root of ESS - log(N * target) instead of log(ESS) - log(N * target), causing incorrect tempering schedules. The fix wraps ess_val in jnp.log() so both sides are in log space:

return jnp.log(ess_val) - target_val

This is consistent with how adaptive_tempered_smc handles the same computation via log_ess() in blackjax.smc.ess.