Resize against a plan baked at translation time, one dispatch per resampled
axis. The plan makes the op independent of the interpolator: nearest, linear
and cubic differ only in window size and weights, so translation is limited
to nodes whose shapes and scales are known then. The scales/sizes input is
kept for arity but no longer read.
Resamples one axis: output[.., x, ..] = sum_k weights[x, k] * input[.., indices[x, k], ..],
with indices already clamped into the axis by the host-built plan.