DEEPFOLD-SOLVER v1.8.9

GPU solver acceleration

• GPU rank-blocker terminal evaluation - the showdown/fold terminal kernel (the dominant ~98% of GPU solve time) is ported from the dense O(nc^2) matrix scan to the O(nc) rank-blocker. The GPU now beats the CPU on flop (~2.7x) and turn (~15x) solves, reversing the prior situation where GPU was slower than CPU.
• Size-aware AUTO backend selection - tiny/river spots run on CPU (avoiding the GPU's fixed per-solve setup cost), while flop/turn solves run on the GPU. AUTO picks the faster backend automatically from tree size and iteration count.
• Validated with a deterministic dense-vs-rank-blocker equivalence self-check (rel diff ~1e-5) across both perspectives, showdown + fold terminals, and single- and multi-runout boards; all parity/correctness tests green.