fix(planner): row handler not included when count(*) is used as subexpression

[xxchan]

fix #501

> explain select count(*) from t;
PhysicalProjection: exprs [InputRef #0]
  PhysicalSimpleAgg: 1 agg calls
    PhysicalTableScan: table #10, columns [], with_row_handler: true, is_sorted: false, expr: None

> explain select 1 from t;
PhysicalProjection: exprs [Int32(1) (const)]
  PhysicalTableScan: table #10, columns [], with_row_handler: false, is_sorted: false, expr: None

> explain select count(*)+1 from t;
PhysicalProjection: exprs [Plus(InputRef #0, Int32(1) (const))]
  PhysicalSimpleAgg: 1 agg calls
    PhysicalTableScan: table #10, columns [], with_row_handler: true, is_sorted: false, expr: None

> explain select -count(*) from t;
PhysicalProjection: exprs [BoundUnaryOp { op: Minus, expr: InputRef #0, return_type: Some(Int(None)) }]
  PhysicalSimpleAgg: 1 agg calls
    PhysicalTableScan: table #10, columns [], with_row_handler: true, is_sorted: false, expr: None

[xxchan]

> explain select count(*)+sum(v) from t
PhysicalProjection: exprs [Plus(InputRef #0, InputRef #1)]
  PhysicalSimpleAgg: 2 agg calls
    PhysicalTableScan: table #10, columns [0], with_row_handler: false, is_sorted: false, expr: None

[skyzh]

Thanks!

Would you please add the cases in issue body to sqllogictests?

By the way, I'm thinking of moving row-handler related logics to the optimizer... For example, if column pruning found that there's no column to scan, we automatically attach a row handler to the table.

At the same time, if our queries become more and more complex, it won't be possible for us to enumerate all rules to add row_handler column, solely in the planner... I don't have a clear idea about that, but I think we will eventually (1) make row_handler a real logical column instead of a flag for table (2) use the optimizer rule to decide whether to scan the row handler column.