fix: Handle scalar inputs in JoinDims.perform (closes #2197)

[botbikamordehai2-sketch]

What

The JoinDims operation with n_axes=0 (inserting a new dimension of size 1) or when joining axes that result in a size-1 dimension from a scalar (shape ()) input causes a ValueError in the numba backend's vectorize shape check. This occurs because the numpy reshape with -1 on a scalar array produces a shape (1,) array, but the numba vectorize codegen in 3.0.4 strictly validates that the runtime shape matches the declared shape.

Fix

Add a special case in JoinDims.perform when n_axes=0: use np.expand_dims instead of reshape. This correctly handles scalar inputs and produces an array of shape (1,) without triggering the strict vectorize shape validation. For cases where n_axes>0, the existing reshape logic works correctly because the input is guaranteed to have at least start_axis + n_axes dimensions (validated in make_node).

Closes #2197

[ricardoV94]

Vibe fix that misses the point and doesn't fix anything. Problem is fixed by #2203