Accelerate `FetchRegionTableIndex` for `TIDB_HOT_REGIONS`

## Enhancement

The current implementation of `FetchRegionTableIndex` is slow:

```
// FetchRegionTableIndex constructs a map that maps a table to its hot region information by the given raw hot RegionMetric metrics.
func (h *Helper) FetchRegionTableIndex(metrics map[uint64]RegionMetric, is infoschema.SchemaAndTable, filter func([]*model.DBInfo) []*model.DBInfo) ([]HotTableIndex, error) {
hotTables := make([]HotTableIndex, 0, len(metrics))
for regionID, regionMetric := range metrics {
...
f := h.FindTableIndexOfRegion(is, hotRange)
...
}

return hotTables, nil
}

// FindTableIndexOfRegion finds what table is involved in this hot region. And constructs the new frame item for future use.
func (*Helper) FindTableIndexOfRegion(is infoschema.SchemaAndTable, hotRange *RegionFrameRange) *FrameItem {
for _, dbInfo := range is.AllSchemas() {
tblInfos, _ := is.SchemaTableInfos(context.Background(), dbInfo.Name)
for _, tbl := range tblInfos {
if f := findRangeInTable(hotRange, dbInfo, tbl); f != nil {
return f
}
}
}
return nil
}
```

It'll fetch and decode all table infos for each region. We can have the following optimization:

1. Only fetch the needed schema for each range.
2. Cache the loaded schema temporarily.

The second optimization is relatively easy. The first one is much more challenging.