model : fix DFlash output head sharing (#30111)
* llama : fix DFlash output head sharing
Assisted-by: Codex
* dflash : read tied output weights from GGUF metadata
Assisted-by: Codex
* llama : share tied word embedding metadata
Assisted-by: Codex
* llama : remove DFlash embedding head fallback
Assisted-by: Codex
