Reject assisted generation for LFM2 and LFM2-MoE (set _is_stateful) (#46937)
LFM2 and LFM2-MoE are conv/attention hybrids that keep recurrent conv state,
but inherited the default _is_stateful = False from Llama. Assisted and
prompt-lookup decoding were therefore not rejected, and since their conv state
cannot be rolled back during speculative verification they silently produced
wrong tokens (divergent from greedy) instead of raising the clear "assisted
generation is not supported with stateful models" error the other stateful
models raise. Set _is_stateful = True on both.