There is a 989.3% performance regression in pxxxxxxr after PR#10621

## Bug Report

### What did you do?

1. Deploy a cluster with 3 TiDB and 3 Tikv.
2. Run benchmark pxxxxxxr.

### What did you expect to see?

No performance regression.

### What did you see instead?

TikvCoprocessorAvg regression in benchmark pxxxxxxr compared to baseline 137.11 after #10621.

| description | tidb | tikv | pd | TikvCoprocessorAvg avg |
| ----------------- | ------------------------------------------------------- | ---------------------------------------------------- | ------------------------------------------------ | ----------------------------------------------------------------------------------------------- |
| last success | pingcap/tidb@84548dbcc1739038be43f4d18de595dc371cea2e | tikv/tikv@b96b66319b5663e2c04e7ed3e9f167bf61ce4ac7 | tikv/pd@7a78ca8644b6adac5f6cb62ccfef897e3cb95156 | 84.67, 109.32 |
| first regression | pingcap/tidb@84548dbcc1739038be43f4d18de595dc371cea2e | tikv/tikv@b96b66319b5663e2c04e7ed3e9f167bf61ce4ac7 | tikv/pd@38a0f9ab0a3e373c2da23fa7c556423b4a67a6db | 1493.55, 1362.70 |

### What version of PD are you using?

tikv/pd@38a0f9ab0a3e373c2da23fa7c556423b4a67a6db

### Root Cause Analysis

PD commit 38a0f9ab (tikv/pd#10621) adds load-based split-scatter. When TiKV reports AskBatchSplit with SplitReason_LOAD, RecordSplitScatterBatch records the batch, and dispatchSplitScatterRegions creates scatter operators on every patrol cycle, causing PD IO doubling, TiKV coprocessor explosion, and extreme tail latency.