Solver refinement

[mapi1]

Adds two NativeSolver knobs and a per-call reduce path for long/large network runs, plus analysis helpers.

New solver kwargs

• grad_horizon: truncated-BPTT gradient horizon. Severs the carry gradient every W steps so the backward pass stops amplifying past one window (the exp(Tλ) blow-up). Forward is bit-identical.
• block_size: unified block unit for rematerialization (checkpointing), per-block fold_in streaming noise, and the reduce grain. Renamed from checkpoint_every. Streaming noise keeps the forward noise tensor at one block instead of O(n_steps).
• reduce= (per-call on solve/prepare): block-level (init, update, finalize) folds that stream a statistic instead of stacking the trajectory, so forward output memory is O(block_size).

Also

• New adiabatic_scan analysis (bifurcation-diagram sweep, expressed as a jitted lax.scan) and refined Lyapunov helpers.
• Noise is now part of the scanned-over axis; solver cleanup/housekeeping.
• Version bump to 0.3.1.