models : consolidate nextn row cropping into shared helpers

[ggerganov]

Overview

• replaces the duplicated nextn output-row crop conditions and the per-model flags (narrow_early, crop_before_ffn, crop_last_layer, emit_h_nextn) with two helpers on llm_graph_context: crop_before_nextn() / crop_after_nextn()
• exactly one of them is true whenever inp_out_ids != nullptr, so the pair documents where the narrowing happens relative to the t_h_nextn capture
• t_h_nextn is now set unconditionally in mimo2, qwen4exp and deepseek4, as in the other nextn-capable models; host reads stay gated by cparams.embeddings_nextn
• the models that only tested embeddings_nextn_masked now share the same condition: with extraction off the last layer is cropped before the capture, which saves its FFN rows - outputs are unchanged
• 20 files (src/llama-graph.h + 19 models)

Additional information

• builds on llama : re-reserve the sched when the nextn extraction flags change #30020: the graph shape still depends on both flags, so the scheduler re-reserves when they change

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES. pi:llama.cpp/MiMo-V2.6-Flash-MOPD

[ServeurpersoCom]

Tested on CUDA: ctest -L main passes, and GLM-4.5-Air greedy outputs are bit-identical to master both without and with draft-mtp (same 59.4% acceptance), with no reallocation under GGML_SCHED_DEBUG_REALLOC=1. The 8 models that now crop late by default show no measurable prefill cost (GLM-4.5-Air and GLM-4.7-Flash, pp512 and pp4096 within noise). LGTM.