feat(v1): add E2B runtime provider

[mattteufel-e2b]

What

Adds an e2b remote sandbox runtime alongside prime and modal (single squashed commit, rebased on current main):

• verifiers/v1/runtimes/e2b.py — the full Runtime contract on the E2B SDK (e2b>=2.35.0): login-shell exec with caller-PATH precedence, durable run_program with stream reconnect (rollouts are never replayed), live processes, background servers exposed via E2B's port proxy, and cancellation-shielded teardown with a sync atexit backstop. Sandboxes boot from a cached template keyed on (image, workdir, cpu, memory); resource constraints are validated up front, including task-level overrides.
• chat.py: drop an explicit tools: null when forwarding chat bodies (some OpenAI-compatible providers reject it; omission is equivalent).
• e2b pytest mark + pairwise e2e placement rows (local-only, like prime/modal), config validation tests, docs.

Open question for maintainers: whether to add an e2b optional extra mirroring verifiers[modal] — left out here to avoid touching pyproject.toml deps; the docs prescribe uv add "e2b>=2.35.0" instead.

Verification (all against live E2B, inference via Fireworks)

• Runtime contract: 31 checks driving E2BRuntime directly — exec/env/cwd/quoting, binary I/O with read caps, run_program, live-process stdio + terminate/kill, run_background + expose reachable from the host, run_uv_script, template cache hit (0.6s second start), idempotent teardown. All passed.
• Edge paths: fresh template build (~150s, then cached), bad image → clean SandboxError, network_access=false blocks egress. All passed.
• Duration/robustness: 12-min run_program and an 8-min-idle open_process both correct; a rollout cancelled mid-program left no sandbox behind (confirmed server-side); 16 concurrent creates all succeeded, paced by the creates_per_sec limiter. All passed.
• e2e matrix: all 7 -m e2b cases green (~48s, -n 4), re-run green after the rebase onto main's limiter + interception-server changes.
• Real eval: gsm8k-v1, bash harness in E2B, 6 tasks x 2 rollouts → 12/12 ok, mean reward 1.0, with the math-verify reward script executing inside each sandbox.
• Hygiene: deterministic suite, ruff, ty green; no leaked sandboxes after any run.

🤖 Generated with Claude Code

[mattteufel-e2b]

Add E2B as a runtime provider