harness: add first-class dataset and ground-truth paths for portable single runs

[jioffe502]

Problem

retriever harness run <benchmark> correctly starts from code-owned benchmark defaults, and --set is useful for expert tuning such as changing query.top_k. However, a developer running one registered benchmark on a workstation without the canonical dataset mount must currently know internal configuration paths:

uv run --frozen --project nemo_retriever retriever harness run jp20_beir \
  --set dataset.path=/raid/data/jp20 \
  --set evaluation.dataset_name=/raid/data/jp20_query_gt.csv

Machine-local corpus and evaluation inputs are normal operator inputs, not experimental benchmark tuning. Operators also should not need to know whether a local queries/qrels source is represented by dataset.query_file or evaluation.dataset_name.

This surfaced during second-host validation of the portable nightly workflow: the nightly dataset map handles host paths, but the direct single-benchmark CLI still exposes internal override keys.

Proposed interface

retriever harness run jp20_beir \
  --dataset-path /raid/data/jp20 \
  --ground-truth /raid/data/jp20_query_gt.csv

For an evaluation-free smoke run:

retriever harness run jp20_smoke \
  --dataset-path /raid/data/jp20

Design notes

• Keep both options optional so existing registry defaults remain unchanged.
• Keep --set as the advanced escape hatch for benchmark tuning.
• Model these as typed run inputs at the benchmark-resolution seam rather than constructing internal --set strings in the Typer command.
• --ground-truth should supply the local queries/qrels input and hide the current dataset/evaluation representation details.
• Define unambiguous precedence. Preferably fail early when a first-class flag conflicts with the corresponding --set key.
• Validate local paths before model, Ray, or GPU initialization.
• Record the final resolved paths in resolved_benchmark.json.
• Remote evaluation datasets such as ViDoRe should remain unchanged.

Acceptance criteria

• retriever harness run jp20_beir continues to use the registered defaults.
• --dataset-path PATH overrides the registered corpus location.
• --ground-truth PATH provides the local queries/qrels source without requiring internal override keys.
• jp20_smoke --dataset-path PATH works without a ground-truth argument.
• Missing or unreadable inputs fail during resolution with actionable messages.
• Conflicting first-class and generic overrides fail clearly.
• resolved_benchmark.json contains the resolved corpus and ground-truth paths.
• Existing --set, runfile, runset, nightly, and remote ViDoRe behavior remains compatible.
• User documentation includes a portable single-benchmark example.

Scope

This makes registered benchmarks portable across machines with different local paths. It does not attempt to infer arbitrary corpus formats: the selected benchmark still defines input type, loader, ingest/query configuration, and evaluation semantics.