Enable NUMA-aware GPU worker binding, config updates after fixes

[karthikeyann]

Optimizes multi-GPU Presto worker performance through NUMA-aware process pinning, cuDF batching tuning, and enabling AUTOMATIC join distribution.

NUMA-aware GPU binding

• Rewrote launch_presto_servers.sh to auto-detect the closest NUMA node per GPU via nvidia-smi topo and pin workers using numactl.
• Installed numactl in the Docker image, granted containers privileged: true, and mounted /sys/devices/system/node for topology visibility.

cuDF performance tuning

Added to config_native.properties:

• cudf.jit_expression_enabled=false — avoids JIT warmup penalty across workers.
• cudf.intra_node_exchange=true — required for UCX to use NVLink.
• cudf.partitioned_output_batch_rows=100000000 — reduces exchange overhead.
• cudf.concat_optimization_enabled=true / cudf.batch_size_min_threshold=100000000 — enables rebatching before aggregations.

Join distribution

• Removed forced PARTITIONED override in generate_presto_config.sh. UCX exchange now supports BROADCAST, so the optimizer can choose automatically.

Bug fix

• Fixed start_presto_helper.sh to only set GPU_WORKER_SERVICE when SINGLE_CONTAINER=false.

Test Plan

• Verify NUMA binding for single-GPU and multi-GPU deployments (check worker logs)
• Run TPC-H/TPC-DS at SF1K+ — no regressions
• Confirm AUTOMATIC join distribution selects BROADCAST where beneficial (EXPLAIN plans)
• Validate batching settings don't cause OOM on large queries