logical_plan optimize for MemTable will cost more time when collect dataframe

### Is your feature request related to a problem or challenge?

I try to read table from MemTable which is large, but when I collect it, datafusion cost more time then expectation.
I find this is because logical_plan optimize.
https://github.com/apache/arrow-datafusion/blob/f9f40bf70a9b6af4345f16b906bcf8a839d5f511/datafusion/core/src/execution/context.rs
```rust
// line 1940
pub async fn create_physical_plan(
&self,
logical_plan: &LogicalPlan,
) -> Result> {
let logical_plan = self.optimize(logical_plan)?;
self.query_planner
.create_physical_plan(&logical_plan, self)
.await
}
```
I find old_plan is TableScan{.., projection: None, ..}, but new_plan is TableScan{.., projection: Some([0,1,2,...]), ..}.
when I collect data, datafusion use new_plan will cost more time then use old_plan because of MemoryStream::poll_next:
https://github.com/apache/arrow-datafusion/blob/f9f40bf70a9b6af4345f16b906bcf8a839d5f511/datafusion/core/src/physical_plan/memory.rs
```rust
// line 203
let batch = match self.projection.as_ref() {
Some(columns) => batch.project(columns)?,
None => batch.clone(),
};
```

### Describe the solution you'd like

_No response_

### Describe alternatives you've considered

_No response_

### Additional context

_No response_

Start in datafusion/core/src/execution/context.rs at create_physical_plan and compare the old and optimized TableScan plans described in the issue. Then inspect datafusion/core/src/physical_plan/memory.rs at MemoryStream::poll_next and measure collection behavior for a large MemTable with and without projection. Done means the optimized path no longer adds the reported collection cost.

Written by the indexing model from the issue text.