Add draft functionality for a generic sandboxed code running

[plaguss]

Description

Work in progress

Add a generic code_runner function to allow running the code in the local environment, or in a sandbox, initially thought for e2b, but should be extendable to other cases.

Examples of calls from LiveCodeBench:

• Call based functions (a function is generated from the LLM)

code = """\
import math
from typing import List

class Solution:
    def maxLength(nums: List[int]) -> int:
        max_len = 1  # At least one element is valid
        n = len(nums)
        for i in range(n):
            current_gcd = nums[i]
            current_lcm = nums[i]
            product = nums[i]
            # Check for single element
            if product == current_gcd * current_lcm:
                if 1 > max_len:
                    max_len = 1
            for j in range(i + 1, n):
                current_gcd = math.gcd(current_gcd, nums[j])
                current_lcm = (current_lcm * nums[j]) // math.gcd(current_lcm, nums[j])
                product *= nums[j]
                # Check if the condition holds
                if product == current_gcd * current_lcm:
                    current_length = j - i + 1
                    if current_length > max_len:
                        max_len = current_length
        return max_len
"""

verification_info = {'test_cases': [{'input': '[1, 2, 1, 2, 1, 1, 1]', 'output': '5', 'testtype': 'stdin'},
 {'input': '[2, 3, 4, 5, 6]', 'output': '3', 'testtype': 'stdin'},
 {'input': '[6,10]',
  'output': '2',
  'testtype': 'stdin'}], 'language': 'python'}

from lighteval.utils.code.runner import code_runner

test_cases = [test["input"] for test in verification_info["test_cases"]]
code = [code]
result = code_runner(test_cases, code, sandbox="e2b", template="call_based", fn_name="maxLength")

• stdio based code (a script that prints to the console)
  ...

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.