Fix stale tuning claim; add attention tags and probe docs

Add Triton flash attention (2-3x over AOTriton, 31x less peak memory)

Fused RMSNorm + GEGLU/SwiGLU Triton kernels for gfx1151

initial commit