Keep fused Qwen prefill on the one-shot lane

MLX 0.32.2 no longer materializes the quadratic D=256 attention score
buffer when the fused Qwen path is active. Skip the stale hybrid OOM
escape hatch for that exact runtime and geometry so affine projections and
GDN state retain the shipped one-shot numerics. Other families, older MLX,
and the explicit fused opt-out retain the existing guard.