feat(wp): tag routed-expert tensors with is_expert at load

Phase 1 Task 1 of the MoE resident/paged split. Add is_expert to
llama_weight_page_info, set from the existing is_consolidated detection
(ffn_*_exps name match) at load. Lets a later filter page only expert
tensors and leave dense weights resident. No behavior change yet.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014ZRfPpL8XFzk1hep9MMg9P

Author: kmbandy

src/llama-model-loader.h

@@ -41,6 +41,10 @@ struct llama_weight_page_info {
     // pager may split it into per-expert sub-pages for routing-aware paging
     // (MAD-88 Phase 2). Set to 1 for non-MoE or per-expert tensors.
     int  n_experts = 1;
+
+    // True iff this is a routed-expert consolidated tensor; dense tensors
+    // (attention, shared expert, embeddings, norms) stay false.
+    bool is_expert = false;
 };
 
 struct llama_model_loader {

src/llama-model.cpp

@@ -1795,6 +1795,7 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
                             info.n_experts = n_exp;
                         }
                     }
+                    info.is_expert = is_consolidated;
                 }
                 ml.weight_page_infos.push_back(info);
                 // Collect the actual model tensor pointer for the weight pager