feat: add Daytona sandbox backend for code execution

[mu-hashmi]

fixes #459

Summary

Adds daytona as a new cloud sandbox backend. This uses the daytona package and the daytona.Daytona sdk.

What changed

• adds rllm/tools/code_tools/daytona_tool.py
• wires PythonInterpreter(backend="daytona") to the new backend
• adds daytona to the [code-tools] extra
• updates the installation docs to mention daytona

Implementation Notes

• uses sandbox.process.code_run(...) instead of the code interpreter/session path, so execution stays stateless per call and matches current CodeTool usage in rllm
• passes through the daytona config knobs that matter here: api_url, snapshot, and env_vars
• retry handling: auth errors fail fast, other errors retry
• maps successful runs to stdout, and non-zero exits to stderr plus the last traceback line in error

Validation

$ python -m rllm.tools.code_tools.daytona_tool
CodeToolOutput(name='daytona_python',
               output=None,
               error=None,
               metadata=None,
               stdout='Hello, world!\n',
               stderr=None)
CodeToolOutput(name='daytona_python',
               output=None,
               error='ZeroDivisionError: division by zero',
               metadata=None,
               stdout=None,
               stderr='Traceback (most recent call last):\n'
                      '  File "<stdin>", line 1, in <module>\n'
                      'ZeroDivisionError: division by zero\n')

[JiwaniZakir]

In daytona_tool.py, the retry logic in forward has a subtle issue: when all retries are exhausted, _restart_sandbox is called to create a fresh sandbox, but then an error is immediately returned without attempting to execute code on the new sandbox — so the restart effort is entirely wasted for the current call. The caller receives an error and must retry manually, defeating the purpose of the restart.

The __main__ block's async test is also broken: interpreter(code=..., use_async=True) passes use_async through **kwargs to forward, which never inspects it, so the call returns a CodeToolOutput synchronously rather than a coroutine — await coro will then raise a TypeError. Either forward needs an async def variant or the example should be removed.

Minor: super().__init__() is called after self._init_sandbox() in __init__, which means sandboxes are already allocated before the parent class is fully initialized — if the base CodeTool.__init__ raises, you'd leak sandbox resources. Calling super().__init__() first would be safer.

[mu-hashmi]

@JiwaniZakir updated PR addressing review comments (apologies for letting the PR go stale for this long!)

[mu-hashmi]

@JiwaniZakir just wanted to gently bump this, updated the PR addressing the issues you brought up. appreciate the review!