Avoid NaN gradients for balanced routing losses

[sylvesterkaczmarek]

The squared coefficient-of-variation losses differentiate through std, producing NaN gradients at perfectly balanced expert loads. This also affects a zero-initialized router despite its finite loss.

Compute the variance directly in the importance and load losses. Cover balanced gates, one-expert routing and router parameter gradients in eager and JIT execution.

Validation: 76 routing, ensemble-routing and MoE tests pass on four virtual CPU devices (two device-count skips). All 22 new regression cases fail on unchanged upstream. Sixty non-balanced value/gradient comparisons match upstream. Syntax and git diff --check pass.