Add fuzzing infrastructure for distributed DataFusion

This change introduces a new `fuzz` binary that enables fuzzing of distributed
DataFusion against a localhost cluster. The fuzzing framework is designed to
be extensible, allowing us to define custom workloads and oracles.

This change defines a TPC-DS workload which uses the duckdb data generator and 99
TPC-DS queries from their github repo. It also defines one oracle, the `SingleNodeOracle`
which validates the set of rows produced by distributed datafusion against the set of
rows produced by single node datafusion (if a query errors in both, then this is not
counted as a failure). You can run this using the following command:
```
RUST_LOG=info cargo run --features integration --bin fuzz -- tpcds --force-regenerate
```

Fuzzing also uses randomized cluster configurations using a deterministic seed.

Next steps:
- Add ordering oracle to validate ORDER BY correctness
  - Idea: Inspect the ordering properties in the logical plan and assert this
    on the RecordBatches
- Observability
  - Log stats on queries that were invalid (ie. failed to execute on single node df)
    so we can measure the quality of queries in a workload
- Add metrics oracle to validate output_rows metric (ensuring metrics are
  working correctly)
- Set up nightly github actions workflow to run the fuzzer automatically
  - Ensure that the data is available to be downloaded so we can reproduce any failures
    locally
- Add SQLancer workload
  - SQLancer produces INSERT and SELECT statements which we could point at a datafusion
    distributed cluster and verify using our oracles
  - Although it doesn't support nested select statements, I was able to get a ~30% error
    rate (meaning 70% of queries were valid datafusion queries)

Author: jayshrivastava

.gitignore

@@ -1,4 +1,5 @@
 /.idea
 /target
 /benchmarks/data/
-testdata/tpch/data/
+testdata/tpch/data/
+testdata/tpcds/data/