DataFrame::cache errors with `Plan("Mismatch between schema and batches")` but query works when not cached

[tv42]

Describe the bug

Dataframe::cache gives an error where an execution that doesn't first cache results succeeds.

I would have expected caching to have no effect on success/failure.

To Reproduce

use datafusion::prelude::SessionContext;

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    let sql = "SELECT CASE WHEN true THEN NULL ELSE 1 END;";
    let ctx = SessionContext::new();
    let plan = ctx.state().create_logical_plan(sql).await?;
    let df = ctx.execute_logical_plan(plan).await?;
    // Comment out the next line to make the error go away.
    let df = df.cache().await?;
    let batches = df.collect().await?;
    let display = datafusion::arrow::util::pretty::pretty_format_batches(&batches).unwrap();
    println!("{}", display);
    Ok(())
}

Expected behavior

Behavior with and without let df = df.cache().await? to be functionally same, only changing performance and memory use.

Additional context

No response

[Asura7969]

The reason is that schema comparison：
https://github.com/apache/arrow-datafusion/blob/d091b55be6a4ce552023ef162b5d081136d3ff6d/datafusion/core/src/datasource/memory.rs#L68

schema：

Field { name: "CASE WHEN Boolean(true) THEN NULL ELSE Int64(1) END", data_type: Null, nullable: true, dict_id: 0, dict_is_ordered: false, metadata: {} }

batches_schema：

Field { name: "CASE WHEN Boolean(true) THEN NULL ELSE Int64(1) END", data_type: Int64, nullable: true, dict_id: 0, dict_is_ordered: false, metadata: {} }

data_type：Null & Int64

Can we use the optimized schema to compare，like：
before：

    pub async fn cache(self) -> Result<DataFrame> {
        let context = SessionContext::new_with_state(self.session_state.clone());
        let mem_table = MemTable::try_new(
            SchemaRef::from(self.schema().clone()),
            self.collect_partitioned().await?,
        )?;

        context.read_table(Arc::new(mem_table))
    }

after:

    pub async fn cache(self) -> Result<DataFrame> {
        let context = SessionContext::new_with_state(self.session_state.clone());
        let physical_plan = self.create_physical_plan().await?;
        let mem_table = MemTable::try_new(
            physical_plan.schema(),
            self.collect_partitioned().await?,
        )?;

        context.read_table(Arc::new(mem_table))
    }

[alamb]

In general, DataType::Null is normally resolved (via Coercion) to an actual type of the target schema

Maybe we could apply the same approach here