Add normalizer option to remove global mean surface temperature

[mcgibbon]

Atmospheric warming/cooling signals show up as large temperature anomalies across many vertical levels at once. Normalizing each temperature field by its global mean leaves a slowly-varying global temperature offset embedded in every channel, which the network has to re-learn at every layer. This PR adds an opt-in normalization mode that strips the global-mean surface temperature anomaly from all temperature fields before normalization (and restores it after denormalization), so the network sees temperature gradients rather than absolute global temperatures.

The offset is computed per sample as surface_temperature_norm_mean - sample.surface_temperature.mean(spatial dims) and added to every temperature field on normalize(), then subtracted back on denormalize(). Cross-channel and horizontal gradients are preserved; only the global mean shifts. The flag is meaningful only for the network normalizer (paired normalize/denormalize); the loss normalizer always strips it because offsets computed independently from target vs. prediction inputs would not cancel cleanly.

Changes:

• fme.core.normalizer.NormalizationConfig.remove_global_mean_surface_temperature: new opt-in bool flag (default False).

• fme.core.normalizer.StandardNormalizer: caches the per-sample offset during normalize() and inverts it in denormalize(); raises if surface_temperature is missing from means or input tensors, or if denormalize() runs before normalize().

• fme.core.normalizer.NetworkAndLossNormalizationConfig.get_loss_normalizer: strips the flag from the loss normalizer via the new _disable_global_mean_surface_temperature_removal helper.

• fme.core.normalizer._GLOBAL_MEAN_SURFACE_TEMPERATURE_REFERENCE / _TEMPERATURE_FIELD_NAMES: hard-coded reference field and target field list.

• StandardNormalizer.from_state reads the legacy use_shared_temperature_offset state-dict key as a fallback so existing checkpoints continue to load.

• Tests added

• If dependencies changed, "deps only" image rebuilt and "latest_deps_only_image.txt" file updated

[mcgibbon]

I'm still not sure the normalization is the right place for the code, as I'm wondering if I should add the global mean surface temperature back in as a separate (constant) field, and the normalizer wouldn't be the right level to do that.

[mcgibbon]

Superceded by #1193.