[feature request] set default cache policy

[kaifronsdal]

I love being able to cache model responses in inspect and the fine grained control I have over the cache policy. However, I often find myself writting a new solver and forgetting to add a cache policy when I call generate or wanting to change the cache policy in a bunch of places all at once. Is there some way to set a defeault cache policy for an eval?

If not, here are some potential options for an interface:

1. In the Task

Task(
    dataset=dataset,
    solver=solver,
    scorer=[...],
    config=GenerateConfig(cache=CachePolicy("1M")), # in the generate config
    cache=CachePolicy("1M") # or a seperate parameter
)

2. As an environment variable

export INSPECT_DEFAULT_CACHE_POLICY="expiry=1M, per_epoch=True"

3. As a argument to inspect eval

inspect eval task.py --model openai/gpt-4o --cache "expiry=1M, per_epoch=True"

[jjallaire]

If we added it to GenerateConfig then all of the above would actually work! (syntax for the CLI and environment variables would be JSON encoded rather than what you show here). We'll plan on doing that.