TransactionContext.UpdateDeltaForTable is inefficient

[ekexium]

Enhancement

In a microbenchmark of AddRecord (with pipelined DML, ART memdb and #53605), UpdateDeltaForTable uses noticeable CPU time.

The primary time consumption is for map iteration, which could be optimized.

ROUTINE ======================== github.com/pingcap/tidb/pkg/sessionctx/variable.(*TransactionContext).UpdateDeltaForTable in /home/eke/code/tidb/pkg/sessionctx/variable/session.go
     180ms      1.38s (flat, cum)  5.90% of Total
         .          .    309:func (tc *TransactionContext) UpdateDeltaForTable(physicalTableID int64, delta int64, count int64, colSize map[int64]int64) {
      10ms       10ms    310:	tc.tdmLock.Lock()
      40ms       40ms    311:	defer tc.tdmLock.Unlock()
         .          .    312:	if tc.TableDeltaMap == nil {
         .          .    313:		tc.TableDeltaMap = make(map[int64]TableDelta)
         .          .    314:	}
      20ms       30ms    315:	item := tc.TableDeltaMap[physicalTableID]
      30ms       30ms    316:	if item.ColSize == nil && colSize != nil {
         .       10ms    317:		item.ColSize = make(map[int64]int64, len(colSize))
         .          .    318:	}
         .          .    319:	item.Delta += delta
         .          .    320:	item.Count += count
         .          .    321:	item.TableID = physicalTableID
      50ms      990ms    322:	for key, val := range colSize {
         .      160ms    323:		item.ColSize[key] += val
         .          .    324:	}
      20ms       90ms    325:	tc.TableDeltaMap[physicalTableID] = item
      10ms       20ms    326:}
         .          .    327:
         .          .    328:// GetKeyInPessimisticLockCache gets a key in pessimistic lock cache.
         .          .    329:func (tc *TransactionContext) GetKeyInPessimisticLockCache(key kv.Key) (val []byte, ok bool) {
         .          .    330:	if tc.pessimisticLockCache == nil && tc.CurrentStmtPessimisticLockCache == nil {
         .          .    331:		return nil, false