Add environment parameter to preserve env vars in eval-retry (#3312)

[evcyen]

This PR contains:

• New features
• Changes to dev-tools e.g. CI config / github tooling
• Docs
• Bug fixes
• Code refactor

What is the current behavior?

Fixes #3312. Environment variables set during an eval run are not recorded in the eval log. When inspect eval-retry replays a failed eval, it reconstructs the task, model config, sandbox settings, and other parameters from EvalSpec, but environment variables are lost because they were never serialized.

This causes eval-retry to behave differently from the original run when the eval depends on environment variables. For example, if an eval uses os.environ.get("MY_CUSTOM_SETTING") to configure sandbox behavior, the retry will fail or behave incorrectly when that variable is not present in the environment.

There is currently no way to pass environment variables through eval() or eval_set() such that they are persisted in the log and recovered on retry.

What is the new behavior?

Adds an explicit environment parameter to the eval() function that allows users to specify which environment variables should be preserved for retry.

Usage:

import os
from inspect_ai import eval, eval_retry

log = eval(
    my_task(),
    model="openai/gpt-4",
    environment={
        "MY_CUSTOM_SETTING": os.environ["MY_CUSTOM_SETTING"],
        "ANOTHER_VAR": "some_value"
    }
)

Does this PR introduce a breaking change?

No. The environment parameter is optional and defaults to None.