Fused indexer top_k (CPU only) (#2098)

* Indexer topk op - CPU only

We do save memory, but it is somehow much slower than what we have on
main.

* Make it a command line option

include/llama.h

@@ -492,6 +492,7 @@ extern "C" {
         bool rope_cache;        // whether to use RoPE cache [EXPERIMENTAL]
         bool graph_reuse;       // whether to reuse graphs when possible [EXPERIMENTAL]
         bool dsa;               // enable GLM DSA sparse attention (off by default) [EXPERIMENTAL]
+        bool fused_idx_topk;    // enable the fused indexer topk op (off by default) [EXPERIMENTAL]
         int  dsa_top_k;         // DSA top-k override (<0 => model's configured indexer_top_k) [EXPERIMENTAL]
         int  min_experts;
         float thresh_experts;