fix(tanh_normal): keep TanhNormal finite on saturated samples

[mctigger]

float32 tanh rounds to exactly ±1 once |x| exceeds about 9, which a Normal with a scale of a few units samples routinely. ClampedTanhTransform fed such values through a bare atanh (±inf) and a log(1 - y² + 1e-6) log-det (floored at log 1e-6 with a 1/eps derivative), so log_prob became -inf and the Monte-Carlo entropy/mode of the SamplingDistribution wrapper went non-finite with NaN gradients. A TanhNormal actor head initialised at std ≈ 5 hit this on every state.

Follow DreamerV2's TanhBijector:

• _inverse clamps y to the largest magnitude strictly below 1 that its dtype represents before atanh, so x stays finite and the gradient through a saturated sample is zero instead of infinite;
• log_abs_det_jacobian uses 2 (log 2 - x - softplus(-2x)) evaluated on x (the same identity torch's TanhTransform uses), exact for every finite x.

Off saturation the density is unchanged (matches the analytic form to ~1e-6). Adds transform-level and distribution-level regression tests for the saturated regime, including half precision.