server: continue GDN prompt cache across retokenized history

Align follow-up tokens to the cached text prefix, then keep the
generated KV and append only the new last-user turn when BPE/jinja
replay would otherwise drop a long assistant prefix.

Assisted-by: Grok