Multi-operand GPU kernels
The sweep grew from unary to any arity (D134). fma, pow, min, max and
division now dispatch over the cross-product of an edge operand set
(+/-0, 1, 3, +/-inf, NaN) - 49 pairs each for binary, 343 triples for
fma - emitting a new gpuop record that carries the output
and each input without disturbing the unary gpu format. 22
kernels, 539 multi-operand observations a run. Arity is a directive in
the shader, so a new multi-operand kernel is still a .comp file plus a
regenerate. verify: ok.