Fix experiment-dataset linking when running evals with a dataset

[alfakini]

The issue

When running evals with a dataset via Braintrust::Eval.run, the resulting experiment is not linked to the dataset in the Braintrust UI. The UI shows "Rows not attached to a dataset" because dataset_id and dataset_version are never included in the experiment creation request:

The issue is that Eval.resolve_dataset resolved a dataset to an array of cases but discarded the Dataset object, so dataset_obj.id was never captured. Experiments#create did not accept or send dataset_id and dataset_version in the POST /v1/experiment payload, even though the API supports both fields.

Additionally, Dataset#version returns nil when the dataset is not explicitly pinned to a version. The Python SDK handles this by computing the version as max(_xact_id) across all records in the fetched dataset, but the Ruby SDK does not.

Fix

Eval.resolve_dataset now returns a hash with :cases, :dataset_id, and :dataset_version instead of a plain array. When no pinned version is available, it computes dataset_version from max(_xact_id) across fetched records (matching the Python SDK behavior).

Eval.run retrieves dataset_id and dataset_version from the resolved dataset and forwards them to the experiment-creation process. Experiments#create accepts optional dataset_id and dataset_version keyword arguments and includes them in the API payload when present.

After the fix got applied:

Tests

Added assertion to the existing dataset eval test verifying that dataset_id and dataset_version are sent in the POST /v1/experiment request body.

Added test_eval_run_without_dataset_does_not_send_dataset_fields to verify that dataset_id and dataset_version are nil when no dataset is provided.

[delner]

Looks good from what I can see! The version lookup is a bit quirky but I think we'll need to rework how datasets are constructed soon anyways (outside the scope of this PR.)

I'll enable the workflow, and if everything is good in CI, we'll merge.

[delner]

I think tests are failing because this PR is from a fork and secrets are not available. I've made #105 to help unblock this. @alfakini you'll want to rebase on this (or on main once it's merged.)

[delner]

This was released in v0.1.4.