Update DFlash recommendation (#4)

Promote RC2->main: INT4/BF16 hybrid release (1M context); card cleanup (context 1M, sampling defers to generation_config, new-checkpoint note)

int4: spinquantless routed-only weights, 256K

Enable thinking by default, preserve reasoning; drop max_new_tokens cap

Mark </assistant> (token 24) as special in tokenizer.json

Mark </assistant> (token 24) as special to match internal serving

Laguna S 2.1 initial release