Fix matmul type_inference for 1-D inputs

[adityasingh2400]

Summary

matmul.type_inference in iOS15/linear.py does not match numpy.matmul (and the op's own value_inference) when either input is 1-D. This is issue #2263.

For the reported case, x.shape == (4,), y.shape == (10, 4, 3):

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

@mb.program(input_specs=[mb.TensorSpec(shape=(4,)), mb.TensorSpec(shape=(10, 4, 3))],
            opset_version=ct.target.iOS15)
def prog(x, y):
    return mb.matmul(x=x, y=y)
# inferred output shape was (1, 3); numpy gives (10, 3)

Root cause

Two separate problems:

• 1-D x: the code promotes x to [1, K], computes the broadcast result, then returns ret_shape[1:] to drop the inserted dimension. The inserted 1 lands at index -2 of the result (the result is [*batch, 1, N]), not index 0, so the slice drops the leading batch dimension instead. (4,) x (10, 4, 3) returned (1, 3) rather than (10, 3).

• 1-D y: never handled. The contraction check reads y_shape[-2] before any promotion, so any 1-D y (e.g. (10, 3, 4) x (4,)) raised IndexError, and the both-1-D scalar case crashed the same way. This is the second symptom reported in the issue thread.

Fix

Follow numpy semantics directly: prepend a 1 to a 1-D x, append a 1 to a 1-D y, run the existing batch broadcast, then remove exactly the inserted axes (x's prepended dim at index -2, y's appended dim at index -1, dropping the higher index first so the lower one stays valid). transpose_y on a 1-D y is now rejected to mirror the existing transpose_x guard. 2-D and batched cases are untouched.

The inferred shape now matches numpy.matmul for every combination of {x 1-D, y 1-D, both, neither} x {transpose_x, transpose_y}, including the scalar both-1-D case and symbolic batch dims.

Test

Added TestMatMul::test_builder_1d_input, a parametrized regression test in iOS14/test_linear.py covering 1-D x, 1-D y, both 1-D, and the transpose variants. It fails on the current code (wrong shapes and IndexErrors) and passes with the fix. Existing matmul tests are unaffected.

Fixes #2263