Reduce pressure for asking gc safe point from PD

[JaySon-Huang]

Each time a TxnMergeTree merge task starts, it will ask PD for gc safe point. If there are hundreds of tables of TxnMergeTree, it may cause PD busy.

Add cached for gc safe point to reduce the num of request to PD.

[solotzg]

We can not control our own gc safe point, because the region will be send to flash by snapshot, which might have already been gc by newest safe point.

[JaySon-Huang]

@solotzg
Do you mean while doing applySnapshot, we should get safe point without cached? If so, I will change it later.

But I still think that while doing merge in TxnMergeTree, we should get safe point with cached, or PD may get busy if there are hundreds of tables in TiFlash. Do you agree?

[JaySon-Huang]

We will only get safe point with cached in Engine DeltaMerge, close this PR.