fuse_reduce_mean: do not fuse when the scalar operand raises the rank

[LeSingh1]

Problem

fuse_reduce_mean's docstring says the multiplier is a const (scalar). The code validates it with _check_var_scalar_value, which accepts any tensor whose size == 1, regardless of rank:

if isinstance(x.val, np.ndarray):
    if x.val.size != 1:
        return False

A shape (1, 1, 1) const holding 1/count therefore passes the check. But mul broadcasts reduce_sum's output up to rank 3, and the reduce_mean that replaces the pair does not:

@mb.program(input_specs=[mb.TensorSpec(shape=(2, 3))])
def prog(x):
    s = mb.reduce_sum(x=x, axes=[1], keep_dims=False)          # (2,)
    return mb.mul(x=s, y=np.array([[[1.0 / 3]]], dtype=np.float32))

PASS_REGISTRY["common::fuse_reduce_mean"](prog)
# ['reduce_sum', 'mul'] -> ['reduce_mean']
# output shape (1, 1, 2) -> (2,)

The real_div branch has the same problem with a shape (1, 1, 1) count. Both are the model's declared output, so the converted model has a different output shape than the program it was built from.

common::fuse_reduce_mean is in the default pipeline.

Fix

Check that the child op's output shape matches reduce_sum's output shape before rewriting. A size-1 operand that does not raise the rank (a plain scalar, or a shape (1,) const against a rank-1 reduce output) broadcasts to the same shape and keeps fusing, so the pattern the pass actually targets is unaffected.

I deliberately put the check in this pass rather than tightening _check_var_scalar_value, which is shared with several other passes that may legitimately want the looser size-1 test.

Tests

In TestReduceMeanFusion:

• test_invalid_pattern_rank_raising_mul — the repro above. Fails on main.
• test_invalid_pattern_rank_raising_real_div — same through the real_div branch. Fails on main.
• test_valid_pattern_size_one_tensor_multiplier — a shape (1,) multiplier that does not raise the rank must still fuse (guards against over-narrowing). Passes before and after.

The existing test_valid_pattern* / test_invalid_pattern* tests are untouched. I ran the whole class before and after; the failure sets are identical (this environment cannot load CoreML.framework, so assert_model_is_valid fails there either way — the graph-structure assertions preceding it all run and pass).

Note: this touches test_cleanup_passes.py, which my PR #2786 also modifies, in a different class.