Add optional CUDA engines and reproducible full-brain odor and voxel experiments

[freeman-1984-coder]

Adds optional CUDA execution while keeping CPU installation and imports independent of CuPy. The ordinary backend preserves existing dimensionless LIF/checkpoint semantics. A separately versioned experimental mV synaptic engine adds decay, delayed transmission and explicit refractory/input scheduling; it never silently reinterprets existing models or checkpoints.

Includes checksum-pinned full FlyWire loading, exact large-ID mapping, annotated bilateral olfactory inputs, numerical and sensory experiments, and a voxel body with recorded GPU playback. All 139,255 neurons and 16,847,997 aggregate connection rows are retained. The old normalized benchmark cannot propagate sensory-only spikes; the new synaptic model does, but its DNa02 readout remains biased. In the actual voxel experiment the body curves away from food, while the input-silenced control stays stationary. Neither reaches food. Raw results, assumptions and unsuccessful outcomes are published.

Validation:

• Independent Brian2 comparison: four fixtures match spike timing, maximum observed voltage error below 5e-13 mV and delayed checkpoint replay.
• Actual A16-8Q: five new synaptic hardware cases, plus 207 software tests in the final voxel source environment, zero skips.
• New full graph: 200 active CPU/GPU ticks, five whole-state comparison points with zero measured float differences; exact discrete state and 20-tick JSON/RNG replay.
• Five matched 800 ms odor conditions; two 2 s full-brain voxel runs. Each voxel run replays a complete future body frame with exact brain/world/RNG state after JSON restore.
• Linux, macOS and Windows CI; the Godot bridge now consumes bounded request bodies before rejecting browser origins to prevent Windows TCP resets. Clean wheel/sdist checks and CPU quickstart/checkpoint passed without CuPy or Brian2.
• Historical dimensionless whole-graph benchmark retained separately: 9.61x relative speedup on the measured host, not real time. The new 0.1 ms synaptic protocol has different performance and should not inherit that figure.

The voxel speed/turn decoder and ground-body mechanics are explicit untrained engineering mappings, not biological locomotion, banana recognition or reliable navigation. Full-brain entry remains research scripts; the experimental mV engine is not yet a FlyBrain.load() catalog model. Existing CPU examples and download opt-in remain available.

Synaptic source and numerical validation · Full odor evidence · Voxel evidence and reproduction