Fix vLLM harness provider kind precedence

[abhinav-bellapu]

Description

Fixes the provider-kind mismatch described in #123 for Beaker launches that use a harness preset such as default with a model preset whose provider kind is vllm.

The root cause was that HarnessConfig.merge_provider() treated harness provider fields as explicit only when they differed from the default ProviderConfig(). Since vllm_server is also the default provider kind, a named harness preset that explicitly selects vllm_server could still lose that kind when merged with a model preset that selects vllm. Beaker job assembly would then install vLLM for the isolated vllm_server path, while runtime config could select the in-process vllm provider.

This change lets callers preserve the harness provider kind when they know it came from an explicit harness preset, config, or override. Empty/default harness configs still allow the model preset provider kind to win, preserving the existing non-harness behavior.

Reviewer notes:

• The fix is applied at the config-merge boundary so runtime provider selection and Beaker dependency assembly share the same provider kind.
• The new preservation behavior is opt-in for call sites that have already materialized an explicit harness config; plain model-provider merges keep their previous precedence.
• Coverage includes the merge helper, CLI config resolution, and Beaker dependency assembly path that originally exposed the mismatch.
• No user-facing docs were changed because this corrects preset precedence behavior without adding new CLI/config syntax.

Closes #123

Type of Change

• Bug fix (non-breaking change that fixes an issue)
• New feature (non-breaking change that adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation update
• Refactoring (no functional changes)

Testing

• Unit tests pass locally (pytest tests/ --ignore=tests/integration/)
• Integration tests pass (if applicable)
• New tests added for new functionality

Local checks run:

• uv run --no-group vllm ruff format --check src/olmo_eval/harness/config.py src/olmo_eval/cli/run/config.py tests/core/harness/test_harness.py tests/cli/run/__init__.py tests/cli/run/test_config.py tests/cli/beaker/test_launch.py
• uv run --no-group vllm ruff check src/olmo_eval/harness/config.py src/olmo_eval/cli/run/config.py tests/core/harness/test_harness.py tests/cli/run/__init__.py tests/cli/run/test_config.py tests/cli/beaker/test_launch.py
• uv run --no-group vllm pytest tests/core/harness/test_harness.py tests/cli/run/test_config.py tests/cli/beaker/test_launch.py::TestHarnessOverridesProviderDependencies
• uv run --no-group vllm pytest tests/cli tests/core/harness

Checklist

• My code follows the project's style guidelines
• I have performed a self-review of my code
• I have added/updated documentation as needed
• My changes generate no new warnings
• Any dependent changes have been merged and published