Make it possible to push Agg down cross UnionScan

## Enhancement
```

id | estRows | estCost | actRows | task | access object | execution info | operator info | memory | disk
-- | -- | -- | -- | -- | -- | -- | -- | -- | --
HashAgg_7 | 1.00 | 110461651.48 | 1 | root |   | time:1.97s, loops:2, partial_worker:{wall_time:1.97215596s, concurrency:5, task_num:2561, tot_wait:9.668152304s, tot_exec:183.727117ms, tot_time:9.86055611s, max:1.97211872s, p95:1.97211872s}, final_worker:{wall_time:1.972170191s, concurrency:5, task_num:9, tot_wait:20.22µs, tot_exec:2.73µs, tot_time:9.86074141s, max:1.97215273s, p95:1.97215273s} | funcs:count(1)->Column#24 | 622.0 KB | 0 Bytes
└─UnionScan_9 | 2468991.74 | 61179069.06 | 2580537 | root |   | time:1.96s, loops:2562 | eq(obei_common_platform.t_robot_api_info.arrange_id, "MATC-03-XJPP-01-001") | N/A | N/A
└─IndexReader_11 | 2468991.74 | 61179069.06 | 2580536 | root |   | time:53.5ms, loops:2526, cop_task: {num: 84, max: 74.7ms, min: 241.6µs, avg: 20.8ms, p95: 72.3ms, max_proc_keys: 50144, p95_proc_keys: 50144, tot_proc: 1.67s, tot_wait: 2.45ms, copr_cache_hit_ratio: 0.39, build_task_duration: 21.3µs, max_distsql_concurrency: 4}, rpc_info:{Cop:{num_rpc:84, total_time:1.75s}} | index:IndexRangeScan_10 | 8.38 MB | N/A
└─IndexRangeScan_10 | 2468991.74 | 532561840.92 | 2580536 | cop[tikv] | table:t_robot_api_info, index:idx_arrangeId(ARRANGE_ID) | tikv_task:{proc max:70ms, min:0s, avg: 28ms, p80:50ms, p95:70ms, iters:2853, tasks:84}, scan_detail: {total_process_keys: 1616984, total_process_keys_size: 163315384, total_keys: 1617035, get_snapshot_time: 998.4µs, rocksdb: {delete_skipped_count: 1311, key_skipped_count: 2687466, block: {cache_hit_count: 6078}}}, time_detail: {total_process_time: 1.67s, total_suspend_time: 2.73ms, total_wait_time: 2.45ms, total_kv_read_wall_time: 1.62s, tikv_wall_time: 1.68s} | range:["MATC-03-XJPP-01-001","MATC-03-XJPP-01-001"], keep order:false | N/A | N/A

```

As you can see, when there're dirty writes, the `HashAgg` can not be pushed down to the coprocessor side.
So the computation is left to the TiDB side.
Compared with the normal execution without dirty writes, it's much slower.