Bound gather_nd / scatter_nd indices by the dims they actually address

[LeSingh1]

iOS17 gather_nd and scatter_nd validate indices against the full input shape:

upper_bound = self.x.shape
if np.count_nonzero(np.logical_or(indices < 0, indices >= upper_bound)):

indices has shape (*K) and its last axis addresses only indices.shape[-1] dims, starting after batch_dims for gather_nd. Comparing a (..., K) array against a length rank(x) tuple only lines up when K == rank(x).

Two in-bounds programs that fail to convert today:

# x is rank 3, index depth is 2
mb.gather_nd(x=np.arange(24, dtype=np.float32).reshape(2, 3, 4),
             indices=np.array([[0, 1], [1, 2]], dtype=np.int32),
             validate_indices=True)
# ValueError: operands could not be broadcast together with shapes (2,2) (3,)

# batch_dims=1, so index 2 addresses x.shape[1], which is 3 long
mb.gather_nd(x=np.array([[1, 2, 3], [4, 5, 6]], dtype=np.float32),
             indices=np.array([[1], [2]], dtype=np.int32),
             batch_dims=1, validate_indices=True)
# IndexError: Indices is out of bounds ... Expected indices between [0, (2, 3))

The second is the K == 1 case, where the comparison broadcasts and then checks every index against every dim.

Fix: slice the shape down to the dims the indices address — data.shape[:K] for scatter_nd, x.shape[batch_dims : batch_dims + K] for gather_nd. Symbolic dims are skipped, since they cannot be compared. Out-of-bounds indices are still rejected, now against the right bound.

Three new tests in TestScatterNd and TestGatherNd, all failing on main. The existing validate_indices tests are unchanged and still pass.