Add multiple code languages

[enaix]

This PR adds R and Julia support to the grader.

• Add R install to environments/runtime/Dockerfile and python/Dockerfile
• Add Julia install to the same dockerfiles
• Add essential R packages (especially to the manual grader)
• Add essential Julia packages (especially to the manual grader)
• Add R code support to the code grader
• Add Julia code support to the code grader
• Modify net policies to allow R/Julia packages

Next steps:

• Add grader-side python isolation (R)
• Add grader-side python isolation (Julia)
• Make sure that data conversion is correct (especially for competitions with advanced pandas features)

Agents TODO:

• Modify ReAct to work with R/Julia
• Add agent-side python binaries isolation and stricter net policies

Limitations:

• Manual grader dockerfile doesn't handle packages installed by the agent
• No R cuda support

[enaix]

Proposed foreignlang python execution prevention method:

• Allow fork/exec syscalls, since it's needed for libraries to function.
• During the agent execution stage, ban any access to python/libpython (by removing the binaries and replacing them with honeypots) plus utilize a stricter net policy / scan downloaded files after each script exec / download.
• During the grader execution, prevent access to the python binary and libpython library, juliacall/pythoncall/pycall/other needs to be handled separately (since libpython is already loaded in memory)

Python scanner

Export check_os_python_access(shell_cmd) function to be used in agent tool calls. Checking the files after each call is not reliable, since a single tool call may immediately download and obfuscate the binary. Alternatively, a better network policy should be used (forbid network access to common python download servers).

Assumption: there is no targeted attack on ml2b to provide obfuscated/encrypted binaries to the model.
Note: honeypot binaries will be used for gathering statistics on unauthorized python call attempts.

[enaix]

We can move julia grading to a separate script, so that it's impossible to access libpython using pythoncall. This will unfortunately add an overhead due to saving to disk + loading the dataset