Use SSHTunnel for slurm-based start_server requests

[Kipok]

Currently to run requests to a model hosted on slurm, we need to do something like this

from nemo_skills.inference.server.code_execution_model import get_code_execution_model
from nemo_skills.code_execution.sandbox import get_sandbox
from nemo_skills.prompt.utils import Prompt, get_prompt
import os

os.environ['NEMO_SKILLS_SSH_KEY_PATH'] = '<path to key>'
os.environ['NEMO_SKILLS_SSH_SERVER'] = '<cluster hostname>'

sandbox = get_sandbox(
    sandbox_type="local",
    host='<cluster host>',
)

llm = get_code_execution_model(
    server_type='trtllm',
    host='<cluster host>',
    sandbox=sandbox,
)

prompt = get_prompt('generic/default', 'llama3-instruct')
prompt.config.system = "Environment: ipython\n\nUse Python to solve user's request."

prompts = [prompt.fill({
    'question': "Calculate how many 'r' are in strawberry.",
})]

extra_generate_params = {
    "code_begin": prompt.config.template.code_begin,
    "code_end": prompt.config.template.code_end,
    "code_output_begin": prompt.config.template.code_output_begin,
    "code_output_end": prompt.config.template.code_output_end,
}

print(prompts[0])
outputs = llm.generate(
    prompts=prompts,
    stop_phrases=prompt.stop_phrases,
    tokens_to_generate=128,
    temperature=0.0,
    random_seed=0,
    **extra_generate_params,
)
print(outputs[0]["generation"])

We should add some function that will set up that SSH connection through the nemo-run tunnel. Ideally also avoid using sshtunnel_requests since vllm server doesn't support that