lax.abs: better error for unsigned inputs

[jakevdp]

Fixes #17958 by explicitly not supporting unsigned inputs to lax.abs.

[sharadmv]

Does np.abs handle unsigned ints? I feel like if it does, then our lowering should handle this case and avoid emitting invalid stablehlo and just return its input.

[jakevdp]

Both np.abs and jnp.abs handle unsigned ints, the latter by avoiding binding abs_p. lax.abs does not support uints, but the error is a long dump from in mlir. This PR changes the error to something more reasonable.

[sharadmv]

I guess I'm asking if it should handle it? Pallas users might call lax.abs and we'd want to lower that to Mosaic.

The MLIR lowering can raise an error instead of the lax.abs function itself to avoid the stablehlo error.

[jakevdp]

In most cases, we define dtype rules for lax primitives so that dtype-related errors are caught during abstract evaluation, rather than letting those errors arise from a lower level. So I think this change is consistent with overall design decisions we've made in the rest of lax. What do you think?