Fix crash folding an exclusive cumsum

[winklemad]

Fixes #2759.

Problem

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

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

@mb.program(input_specs=[])
def prog():
    return mb.cumsum(x=np.array([1.0, 2.0, 3.0, 4.0], dtype=np.float32), axis=0, exclusive=True)
# TypeError: Cannot construct a dtype from an array

cumsum.value_inference did:

data = np.concatenate((np.zeros(zero_shape, data)), axis=axis)

which (1) passes the data ndarray as np.zeros' dtype argument → TypeError, and (2) wraps a single array in np.concatenate instead of a (zeros, data) sequence. Beyond the crash, the exclusive prefix sum was never actually computed — the inclusive data was returned unshifted. This is reachable from real frontends (e.g. tf.cumsum(..., exclusive=True) on a foldable tensor).

Fix

Build the leading zeros with data.dtype and shift the inclusive cumsum one step along axis (prepend a zero, drop the last element) to produce the exclusive prefix sum out[0] = 0, out[i] = sum(x[:i]). reverse is unaffected — it still flips around the exclusive step.

Tests

Added test_builder_eval_exclusive, covering the folded exclusive cumsum across every axis and reverse value against the numpy ground truth. It raises TypeError before this change and passes after; the existing TestCumSum cases are unchanged. Verified the value-inference path locally (the @ssa_fn eval tests run without the CoreML runtime).

[TobyRoseman]

CI: https://gitlab.com/coremltools1/coremltools/-/pipelines/2683295847

[TobyRoseman]

Thanks @winklemad