Used to read a prediction back out into the original units — a value head trained against symlog targets predicts in the compressed space, and this is what turns that back into a return.

Uses expm1 rather than exp(·) - 1, which loses precision for small arguments exactly where symlog sends most of its inputs.