Fix zeros_like/ones_like/empty_like for scalar and array_like inputs

[cycsmail]

Currently the *_like helpers assume other_ary is a GPUArray and read .allocator/.flags off it, so passing a scalar or other array_like blows up. This mirrors numpy by running non-GPUArray inputs through np.asarray and falling back to the default drv.mem_alloc allocator; the GPUArray path is unchanged.

Added a small test for scalar / np.float32 inputs alongside the existing *_like checks.

Heads up: I don't have a CUDA box here, so I only exercised the _array_like_helper logic locally and couldn't run the GPU suite. Worth confirming on your GitLab CI.

Fixes #389.