manual repartitioning overridden by physical optimizer

**Describe the bug**

Using dataframe.repartition() function doesn't work as expected.

**To Reproduce**
Using the tpch bin from benchmarks, convert the .tbl (csv) files to Parquet format using the "partitions" option:

```
cargo run --release --bin tpch -- convert --input ./tpch-data --output ./tpch-data-parquet --format parquet --partitions 4
```

That should have produced 4 parquet files per table, but instead created 20 (this laptop has 20 cores).

**Expected behavior**

Expected it to produce the specified number of partitions (4 in this case).

**Additional context**
I added some debug and the physical plan produced is:
```
CoalesceBatchesExec: target_batch_size=4096
RepartitionExec: partitioning=RoundRobinBatch(20)
RepartitionExec: partitioning=RoundRobinBatch(4)
RepartitionExec: partitioning=RoundRobinBatch(20)
CsvExec: files=[home/kmitchener/dev/arrow-datafusion/benchmarks/tpch-data/region.tbl], has_header=false, limit=None, projection=[r_regionkey, r_name, r_comment]
```

The tpch bin repartitions the file using this bit of code:
```rust
// optionally, repartition the file
if opt.partitions > 1 {
csv = csv.repartition(Partitioning::RoundRobinBatch(opt.partitions))?
}
```

Start by running the benchmarks/tpch binary with the reproduction command and inspect the physical plan around dataframe.repartition(). Trace why the physical optimizer adds RoundRobinBatch(20) repartitions after the requested 4-partition repartition. Done means the command produces four Parquet files per table as expected.

Written by the indexing model from the issue text.