## Purpose
Fix compilation error due to missing atomic include
## Modifications 
Add #include <atomic>
## Test

## Purpose
Use get_world_group().rank to mark worker's connector metrics, use
engine_id to mark scheduler's connector metrics.
## Modifications 
When using DP, metrics could not figure out which worker the stats
belong to.

## Purpose
rm unavailable lib and fix doc and update patch for blend
## Modifications 

## Test
DATA_DIR=/home/data/kv_cache
MODEL_PATH=/home/models/mistralai/Mistral-7B-Instruct-v0.2
BLEND_DATASET_PATH=/home/datasets/LongBench/data/2wikimqa.jsonl
cd unified-cache-management
python examples/offline_inference_blend.py
<img width="1261" height="255" alt="image"
src="https://github.com/user-attachments/assets/7ab1de8a-3161-4308-a681-10c299e38c41"
/>

---------

Co-authored-by: wuhuxiao <whu.whx@gmail.com>
Co-authored-by: wenxinwang <wangwenxin21@huawei.com>

…re (#711)

…ref (#707)

Reduce directory lock conflicts during batch dumps in PosixStore,
improve its dump performance.

- Split the unified concurrency configuration item (`stream_number`)
into two independent concurrency configurations
(`posix_lookup_concurrency` and `posix_data_trans_concurrency`),
allowing for individual adjustment of the `Lookup` concurrency.
- The directory creation action during `Dump` is moved to the `Setup`
stage, reducing metadata operations during `Dump`.
- Write multiple files during the `Dump` process to different shard
directories to reduce directory conflicts during concurrent file writes.