Fix MLX rholrate to ratchet at maxdecs (#195)

[neuromechanist]

Closes #195.

Mirrors the torch/numpy fix from #194 for the MLX backend: rholrate is now a maxdecs-ratcheted ceiling, not a per-LL-decrease monotone decay. The old self.rholrate *= self.rholratefact on every decrease collapsed the rho rate toward ~1e-5 and froze the shape.

Design decision (the open question in #195)

Fortran gates the rholrate0 ceiling ratchet on iter > newt_start independent of do_newton (amica15.f90:1049 -- unlike the newtrate ratchet one line below, which also requires Newton). So MLX keeps newt_start (default 50) purely as that schedule threshold, even though it does no Newton. rholrate is reset to rholrate0 at fit start and, since MLX has no rho ramp, nothing re-inflates it -- so it ratchets only at maxdecs.

Tests (real sample EEG, Apple GPU)

Two new tests, both passing on the GPU:

• test_rholrate_ratchets_at_maxdecs_not_per_decrease: an aggressive-lrate run triggers several LL decreases; the surviving rholrate ratcheted at the maxdecs cadence (rholrate0 * factor^k), orders of magnitude above the old per-decrease decay.
• test_rholrate_ceiling_ratchet_gated_on_newt_start: with newt_start past the budget, decreases never move the ceiling (gate holds).

Full MLX suite: 10 passed. ruff + ty clean.