Demonstrate task-attached default run configuration

[MattFisher]

Purpose

Demonstrates what the task-attached default configuration from Inspect AI PR 5294 enables for ExploitBench.

The normal invocation becomes:

inspect eval exploitbench/exploit_bench --model provider/model

Inspect discovers and applies the maintained YAML automatically, while an explicit --run-config still replaces it.

What this changes

• attaches run_configs/default.yaml with @task(default_config=...)
• makes task parameters config-supplied instead of loading YAML into Python defaults
• moves generation, limits, epoch reduction, and error-scoring policy into the YAML profiles
• deletes the production run_config.py loader
• retains package-resource loading only for eval.yaml and k8s.yaml metadata
• updates CLI, Modal, Python API, provenance, and override coverage
• pins the reviewed head of Inspect AI PR 5294 until the feature is released

Why draft

This is a demonstration stacked on codex/exploitbench-review-fixes and depends on the unreleased Inspect AI feature. It is intended to make the resulting API and code shape easy to review before choosing whether to adopt it.

Validation

• 300 passed, 23 skipped
• Ruff formatting and lint pass
• mypy passes
• uv lock check passes

[pabloRom2004]

Will we have to wait here for the unreleased feature to Inspect AI to be merged before we can merge this PR?