fix(nx): propagate NaN through clip on BinaryBackend

[blasphemetheus]

Nx.clip on BinaryBackend returns a different thing for NaN depending on which argument carries it:

iex> Nx.clip(Nx.tensor(:nan), 0.0, 2.0)
#Nx.Tensor<f32 0.0>
iex> Nx.clip(Nx.tensor(1.0), Nx.Constants.nan(), 2.0)
#Nx.Tensor<f32 2.0>
iex> Nx.clip(Nx.tensor(1.0), 0.0, Nx.Constants.nan())
#Nx.Tensor<f32 NaN>

The first case is the harmful one: a NaN input comes out as an in-range value, so a clip used as a guard hides the NaN instead of surfacing it. EXLA returns NaN for all three, which matches min(max(x, lo), hi) with Nx's own NaN-propagating min and max.

The cause is the two-comparison chain in BinaryBackend.clip: every comparison against NaN is false, so each position falls through to a different arm. This replaces the chain with the backend's existing element_max and element_min, which already handle NaN and infinities, so both backends compute the same composition.

The new test covers NaN in each argument position and fails on main with the values above. The existing clip tests, including the infinities case, still pass, and so does the full nx suite. (claude) also checked BinaryBackend against the EXLA CUDA client on the three NaN cases plus ordinary float and integer inputs, and they agree on all of them.

🤖 Generated with Claude Code