Use a closed custom JVP for jax.nn.soft_sign and fix +-inf outputs.
Previously, `jax.nn.soft_sign(x) = x / (|x| + 1)` returned `NaN` at `+-inf` (`inf / inf`) and relied on quotient-rule differentiation `1 / (|x| + 1) - |x| / (|x| + 1)^2`. For `|x| >> 1`, subtracting those two nearly equal terms lost half the precision by `|x| ~ 1 / sqrt(eps)`, collapsed `soft_sign'(x)` and higher derivatives to `0` once `|x| + 1 == |x|`, and produced `NaN` gradients at `+-inf`.
Evaluate `soft_sign` via a `@custom_jvp` helper returning `(y, d) = (soft_sign(x), 1 / (|x| + 1))` with `soft_sign(+-inf) = +-1` and closed recurrence `y' = d^2`, `d' = -sign(x) * d^2`.
PiperOrigin-RevId: 1009858470