Softmax along dim — $\sigma(x)_i = e^{x_i} / \sum_j e^{x_j}$.

Dispatches to the engine's numerically stable softmax kernel (max-subtraction baked in), so no caller-side stabilisation is needed. Equivalent to lucid.nn.functional.softmax but lives in the top-level lucid namespace for parity with NumPy / SciPy-style usage.