Import Into: potential regression after #62419 removed coarse split/scatter for large IMPORT INTO

[OliverS929]

Bug Report

1. Minimal reproduce step (Required)

1. Run IMPORT INTO with a very large dataset (global sort + ingest path).
2. During ingest, the generated split key count is large (len(splitKeys) > 100).
3. On builds that include #62419 (rate limiter for split/ingest), region split/scatter runs in one layer only (coarse layer removed).
4. Under this condition, region distribution becomes highly skewed across stores, and write pressure concentrates on a few stores.
5. TiDB ingest may repeatedly hit retryable TiKV write errors (e.g. needRescan) and loop on retries.

Observed signals in logs/monitoring:

• Frequent retries in needRescan stage.
• Errors around opening/sending write streams to TiKV (Unavailable, connection refused/reset/timeout, EOF).
• Region count distribution skew across stores during ingest.

2. What did you expect to see? (Required)

• For large split key sets, split/scatter should spread pressure early and avoid region concentration.
• Existing limiter behavior from #62419 should still be retained to protect PD/TiKV from excessive split requests.
• IMPORT INTO should not amplify store-level pressure due to missing coarse split/scatter stage.

3. What did you see instead (Required)

• Severe region skew during ingest.
• Repeated retry/rescan loops in ingest write path.
• Store-level pressure concentrates on a subset of stores; manual scatter may mitigate temporarily.

4. What is your TiDB version? (Required)

• Exact build hash needs confirmation.
• Regression suspicion is tied to builds containing #62419:
  • #59609 introduced 2-level split/scatter.
  • #62419 added limiter but removed the coarse split/scatter layer.