Docs: lead with the DCP=4 flagship spec — 250k ctx, ~3x concurrency, ~60 tok/s

DCP=4 is the recipe, not a slow tradeoff. Reframe the DCP table so DCP=4 is the
headline (250k / ~3x / ~60 tok/s warm) and DCP=1 is the optional smaller-ctx
fast-path. Update Performance + intro to ~60 tok/s decode and ~3x (2.84x,
710,593-token pool) concurrency.