fix(whiten): scale-only path when shift_mean is false

[YeonwooSung]

Fixes #1369.

distributed_masked_whiten(..., shift_mean=False) used to center, scale, then add the mean back:

whitened_values = (values - global_mean) * rsqrt(var)
if not shift_mean:
    whitened_values += global_mean

That is not a scale-only path. This change makes shift_mean do what it says:

• shift_mean=True (default): (values - global_mean) * rsqrt(global_var + epsilon) — zero-mean, unit-variance
• shift_mean=False: values * rsqrt(global_var + epsilon) — scale only; do not subtract the mean, and do not add it back

Both paths still share the same Bessel-corrected global variance estimate.

The live GRPO caller in slime/backends/megatron_utils/loss.py already uses shift_mean=True, so the production whitening path is unchanged.

A CPU unit test (NUM_GPUS = 0) mocks torch.distributed.all_reduce as a no-op so it can run without a process group, and pins the two formulas plus the shared variance.