feat: add E2B sandbox support

[alt-glitch]

Summary

Add a new E2BREPL environment that runs Python code in E2B sandboxes, following the same HTTP broker pattern as ModalREPL and PrimeREPL.

Changes

• rlm/environments/e2b_repl.py - Full E2BREPL implementation using HTTP broker pattern
• examples/e2b_repl_example.py - Example demonstrating basic code execution and LLM queries
• rlm/environments/__init__.py - Register "e2b" environment
• rlm/core/types.py - Add "e2b" to EnvironmentType
• pyproject.toml - Add e2b optional dependency (e2b-code-interpreter>=0.0.11)

Key Implementation Details

• Uses e2b-code-interpreter SDK with Sandbox.create() API
• Runs Flask broker on port 8889 (not 8888, which E2B uses for Jupyter)
• Uses sandbox.get_host(port) for public URL access
• Supports llm_query(), llm_query_batched(), FINAL_VAR()
• State persistence across code executions via dill serialization

Usage

# Install with e2b support
uv pip install -e ".[e2b]"

# Set API key
export E2B_API_KEY=your_key

# Run example
python -m examples.e2b_repl_example

Example Output

============================================================
E2B REPL Example
============================================================

[1] Basic code execution (no LLM handler)
----------------------------------------
E2B sandbox created, ID: ikk8xw53nne4bytcdg87t
Executed: x = 1 + 2
Locals: {'x': '3'}
Executed: print(x * 2)
Stdout: 6
FINAL_VAR('answer'): 42

[2] Code execution with LLM handler
----------------------------------------
LM Handler started at ('127.0.0.1', 49536)
E2B sandbox created, ID: i0o2ud044jn49rgy4pvp8
Broker URL: https://8889-i0o2ud044jn49rgy4pvp8.e2b.app
Executed: response = llm_query('What is 2+2?')
Stderr: (none)
Response: Mock response to: What is 2+2?

Executed: responses = llm_query_batched([...])
Result: Got 3 responses
First response: Mock response to: Question 1

============================================================
Done!
============================================================

Test Plan

• Basic code execution works
• State persistence works across calls
• FINAL_VAR() extracts variables correctly
• LLM handler integration works (single and batched queries)
• Broker communication works over E2B public URL
• Ruff passes all checks

[alexzhang13]

I'll take a look at this when I have time, I do want to add E2B support.

[alt-glitch]

@alexzhang13 ready for review.

To test, run:

uv run python examples/e2b_repl_example.py