Add per-sample spectral whitening option to the energy-score loss

[mcgibbon]

Small-scale (high-l) spectral power converges very slowly in SFNO training because the loss lives in a red (grid) spectrum: high-l modes carry tiny amplitude, so they contribute little to the energy score and their dedicated dhconv filter weights see proportionally small, low-SNR gradients. A per-l gradient probe on a trained checkpoint confirmed an ~8× monotone decline of high-l gradient magnitude (and the energy score, though computed per (l,m) mode, uses uniform mode_weights, so it does not correct this).

This PR adds an opt-in spectral_whitening='per_sample' mode to EnergyScoreLoss that reweights the per-(l,m) energy score by the inverse per-degree RMS amplitude of each target sample. This flattens each target's angular power spectrum so high-l modes are no longer starved, raising their gradient SNR. The factor is computed from the detached target coefficients (no new gradient path) and is magnitude-preserving (rescaled per (sample, channel) so the overall energy-score magnitude — and the meaning of energy_score_weight — is unchanged; only the per-scale balance shifts). A white-spectrum target yields a uniform factor (no-op). Default 'none' is bit-for-bit backward compatible.

An A/B validation run (a perturbation of a fg16 sr0.125 residual SFNO baseline, identical except whitening enabled) showed the targeted effect with neutral skill:

• Small-scale spectral fidelity (the target): large win. Train smallest_scale_norm_bias reaches the baseline's final (epoch-120) value by epoch ~11 and finishes ~2× lower (0.031 vs 0.061). Inference-time spectral bias also improves substantially (e.g. 10year h500 1.23 → 0.40, TMP850 0.13 → −0.006).
• Probabilistic skill (CRPS): roughly neutral. Validation 1-step CRPS is ~1% better on every field (geomean 0.989); day-5 CRPS is tied on one weather period (geomean 1.001) and ~2% worse on another (1.019), within single-seed noise. (The training loss value is not comparable across arms because whitening redefines the energy-score term; CRPS is the proper definition-stable comparison.)

Changes:

• fme.core.loss.EnergyScoreLoss — add spectral_whitening / whitening_eps_frac args and _spectral_whitening_factor (per-(sample, channel, l), detached, magnitude-preserving, floored at whitening_eps_frac of the per-sample mean degree amplitude); applied as a multiplicative reweight after mode_weights.

• fme.core.loss.EnsembleLoss — thread energy_score_whitening / energy_score_whitening_eps_frac kwargs through to EnergyScoreLoss.

• fme.core.test_loss — tests for the no-op default, white-target invariance, small-scale boost, magnitude preservation, and config wiring.

• Tests added

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

Resolves # (delete if none)

[mcgibbon]

Make this a Literal[(options)] instead of str

[mcgibbon]

Similar type comment.