cumsum with exclusive=True crashes when constant-folded (TypeError: Cannot construct a dtype from an array) #2759

Description

Constant-folding a cumsum op with exclusive=True crashes at graph-construction time.

Reproduction

from coremltools.converters.mil.mil import Builder as mb
import numpy as np

@mb.program(input_specs=[])
def prog():
    x = np.array([1.0, 2.0, 3.0, 4.0], dtype=np.float32)
    return mb.cumsum(x=x, axis=0, exclusive=True)

File ".../mil/ops/defs/iOS15/tensor_operation.py", line 158, in value_inference
    data = np.concatenate((np.zeros(zero_shape, data)), axis=axis)
TypeError: Cannot construct a dtype from an array

Cause

In cumsum.value_inference, np.zeros(zero_shape, data) passes the data ndarray as np.zeros' dtype argument, and (np.zeros(...)) is a single array rather than a tuple for np.concatenate. So any exclusive=True cumsum on a constant-foldable input aborts conversion — this is reachable from real frontends (e.g. tf.cumsum(..., exclusive=True) on a foldable tensor). Even setting the crash aside, the exclusive prefix sum is never computed: data (the inclusive cumsum) is returned without being shifted.

Expected

Exclusive cumsum of [1, 2, 3, 4] is [0, 1, 3, 6] (out[0] = 0, out[i] = sum(x[:i])).

I'll open a PR that mirrors the exclusive semantics and adds value-inference coverage across axis/reverse combinations.