Fix JAX 0.4.28 regression in SciPy logsumexp scale propagation.

[balancap]

New JAX version introducing an addtional max operation with -inf.
Needs proper handling of this special value in scalify.

[balancap]

Issue due to change in jaxpr: c:f32[] = max -inf b added since JAX 0.4.28

JAX 0.4.28

{ lambda ; a:f32[10]. let
    b:f32[] = reduce_max[axes=(0,)] a
    c:f32[] = max -inf b
    d:bool[] = is_finite c
    e:f32[] = select_n d 0.0 c
    f:f32[] = stop_gradient e
    g:f32[1] = broadcast_in_dim[broadcast_dimensions=() shape=(1,)] f
    h:f32[10] = sub a g
    i:f32[10] = exp h
    j:f32[] = reduce_sum[axes=(0,)] i
    _:f32[] = sign j
    k:f32[] = abs j
    l:f32[] = log k
    m:f32[] = add l f
  in (m,) }

JAX 0.4.26

{ lambda ; a:f32[10]. let
    b:f32[] = reduce_max[axes=(0,)] a
    c:bool[] = is_finite b
    d:f32[] = select_n c 0.0 b
    e:f32[] = stop_gradient d
    f:f32[1] = broadcast_in_dim[broadcast_dimensions=() shape=(1,)] e
    g:f32[10] = sub a f
    h:f32[10] = exp g
    i:f32[] = reduce_sum[axes=(0,)] h
    _:f32[] = sign i
    j:f32[] = abs i
    k:f32[] = log j
    l:f32[] = add k e
  in (l,) }