[BUG] jvp of mx.logcumsumexp is not implemented

Describe the bug

Forward-mode autodiff (mx.jvp) through mx.logcumsumexp raises instead of returning the tangent:

RuntimeError: JVP is not implemented for cumulative prod/min/max

Scan::jvp only implements the Sum reduction; logcumsumexp (the LogAddExp scan) falls through to the generic "not implemented" branch even though its vjp is implemented. This breaks mx.jvp and forward-over-reverse compositions that go through logcumsumexp.

To Reproduce

import mlx.core as mx

x = mx.array([1.0, 2.0, 3.0])
t = mx.ones_like(x)
mx.jvp(lambda z: mx.logcumsumexp(z), [x], [t])
# RuntimeError: JVP is not implemented for cumulative prod/min/max

Expected behavior

mx.jvp of logcumsumexp returns the running softmax-weighted sum of the tangents,

d/dt logcumsumexp(x)_k = sum_{i<=k} softmax(x)_i * t_i,

consistent with the existing vjp (the adjoint identity <w, Jv> == <v, Jᵀw> should hold).

Desktop

• OS: macOS
• Version: main (0.32.0.dev)

Start at Scan::jvp and trace the LogAddExp scan path used by mx.logcumsumexp, comparing it with the existing vjp implementation. Reproduce the Python mx.jvp example, then verify that the tangent is returned and that the stated adjoint identity holds.