Make batched vs list-model training test platform-robust

[gian21391]

test_train_model_batch_compare_adam fails on macOS/ARM: tiny float64 reduction-order differences between the batched and per-model kernels (~1e-10) get amplified by Adam's adaptive normalization, exceeding the 1e-5 tolerance within a few steps.

Rather than loosening the tolerance, this rewrites the test to check the invariant directly:

• Compare loss and per-parameter gradients before any optimizer step (tight, 1e-9) — this is the actual batched-vs-independent equivalence.
• Use plain SGD for the 10-step trajectory comparison; without adaptive normalization the divergence stays at float64 noise level, so the tolerance is tightened from 1e-5 to 1e-8.
• Keep the final lengthscale comparison, tightened to 1e-8.

Measured on the previously failing platform: worst per-step loss difference 9e-16, max gradient difference 6e-17 (vs 2.7e-4 under Adam).

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!