GPU set widened, Deck device selection
Closed the three GPU follow-ups in order. Multi-operand (D134):
fma/pow/min/max/division over an edge cross-product, new
gpuop record. Deck target (D112): device selection prefers
real silicon over llvmpipe, and gpudev carries the device
type so a cpu result is gradable as non-hardware. Widened
set (D113): ten more transcendentals and eight more edge inputs,
including subnormals and range-reduction stressors.
32 kernels now, ~1000 GPU observations a run, all device-labelled. verify: ok throughout, with the shader-drift and GPU=1 compile gates carrying it. The whole set is proven on llvmpipe (software) and is a device swap from RDNA2 truth on a Deck.