Reserve temp_kv bucket space by miss count to avoid rehashing (#5912)

Summary:

X-link: facebookresearch/FBGEMM#2831

Reviewed By: kunalspathak

Differential Revision: D108763406

Author: PerfAICT Bot

fbgemm_gpu/src/dram_kv_embedding_cache/dram_kv_inference_embedding.h

@@ -303,6 +303,7 @@ class DramKVInferenceEmbedding
                         // the blocking time for inference read
 
                         std::unordered_map<int64_t, weight_type*> temp_kv;
+                        temp_kv.reserve(miss_info.size());
                         auto* pool = kv_store_.pool_by(shard_id);
                         auto mem_pool_lock = pool->acquire_lock();
                         for (auto& [id, tensor_offset] : miss_info) {