[bug fix] Fix 25p noise and noise upscaling

[bmhowe23]

Summary

Resolves #20
Resolves #80

This does not affect the published paper results; they report uniform depolarizing noise and were trained via the scalar p_error path, which is unchanged.

Authored by: @kvmto

Changes

• Wire noise_model end-to-end (training/train.py, data/generator_torch.py, qec/precompute_dem.py, qec/surface_code/memory_circuit_torch.py): forward the active 25p model from the trainer to the data generator, DEM precompute, and MemoryCircuitTorch. New p_override lets cached DEM structure be reused with a freshly built probability vector.
• Fix build_single_p_marginal: emit p_prep_X / p_prep_Z for ancilla-prep entries instead of 0.0.
• Fix grouped-totals / upscaling (qec/noise_model.py, fixes Update noise scaling #20): keep P_prep_X/P_prep_Z and P_meas_X/P_meas_Z separate, and compare 0.5 × P_idle_spam_raw. The p = 6e-3 config now sits exactly at target.
• DEM artifact metadata (qec/precompute_dem.py): write versioned metadata + NoiseModel.sha256() into .p.npz; structural mismatch rebuilds in memory, structural match reuses cached H/A and refreshes p. Legacy metadata-free artifacts still load.

Tests

New tests in code/tests/test_noise_model.py and code/tests/test_torch_setup.py cover: canonical sha256 stability, the new 7-channel grouping, the p = 6e-3 at-target case, DEM precompute producing 25p (not scalar-derived) probabilities, metadata round-trip, generator using noise_model for in-memory and on-disk DEM, structural-match cache reuse with refreshed p, and legacy metadata-free artifacts still loading.