Honor interleave in concat's value_inference

[LeSingh1]

Problem

concat.value_inference always calls np.concatenate(values, axis=self.axis.val) and never reads self.interleave. So a concat of const inputs with interleave=True const-folds to the plain, non-interleaved concatenation.

Using the example spelled out in the op's own docstring:

in1 = [[1, 2], [3, 4], [5, 6]]     # (3, 2)
in2 = [[7, 8], [9, 10], [11, 12]]  # (3, 2)

@mb.program(input_specs=[])
def prog():
    return mb.concat(values=[in1, in2], axis=0, interleave=True)

	value
docstring / runtime	[[1,2],[7,8],[3,4],[9,10],[5,6],[11,12]]
value_inference	[[1,2],[3,4],[5,6],[7,8],[9,10],[11,12]]

type_inference computes the correct shape (6, 2), so nothing flags the mismatch, and common::const_elimination then replaces the concat with a const holding the wrong values — the compiled model computes the wrong answer at runtime:

@mb.program(input_specs=[mb.TensorSpec(shape=(6, 2))])
def prog(x):
    c = mb.concat(values=[in1, in2], axis=0, interleave=True)
    return mb.add(x=x, y=c)

apply_pass_and_basic_check(prog, "common::const_elimination")
# ['concat', 'add'] -> ['add'],  with the non-interleaved constant baked in

Fix

All inputs of an interleaved concat share the same shape (type_inference enforces exactly this), so interleaving along axis is a stack immediately after axis followed by folding that new dimension back into axis.

The all-scalar branch above is deliberately left alone: interleaving N scalars and stacking them produce the same result.

I checked the folded value against a slice-assignment reference (out[k::N] = values[k] along the axis) for ranks 1–3, every positive and negative axis, and 2 or 3 inputs — all match.

Tests

In TestConcat (ops/tests/iOS14/test_tensor_operation.py):

• test_builder_eval_interleave[axis=0,1,-1,-2] — checks the folded value against the slice-assignment reference, including negative axes.
• test_builder_eval_interleave_docstring_example — the exact example from the op docstring.

All five fail on main and pass with the fix.

The existing interleave coverage, TestConcat.test_builder_to_backend_stress_interleave in ops/tests/iOS14/test_tensor_transformation.py, feeds mb.placeholder inputs, so value_inference never runs there — which is why this went unnoticed.

I ran all of ops/tests/iOS14/test_tensor_operation.py and the concat-related pass tests before and after; the failure sets are identical (this environment cannot load CoreML.framework, so the run_compare_builder / assert_model_is_valid tests fail there either way).