hybrid: prefill cached follow-ups above 4*decode

Share the first-prompt 4*decode threshold so a 768-token chat leftover
uses ubatch 2048 instead of decode 64. Tiny leftovers stay on decode
graphs. Rewrite start-ling-tiny.sh onto the start1660 knob surface with
ngram n=12 m=4 (disables 12-48 token verifies) and KVFlash 8192.

Assisted-by: Grok