Move shared Rich console to olmo_eval.common.console to break import cycle

[finbarrtimbers]

Moves the shared Rich console singleton from olmo_eval.cli.utils to a new olmo_eval.common.console module so non-CLI modules don't pull the entire cli package into their import graph.

The cycle this prevents

When a module under olmo_eval.harness.sandbox is imported at module level — as finbarr/prebuild-sandbox (#159) does in scicode/eval.py — the chain is:

from olmo_eval.harness.sandbox import SandboxManager
  → harness/__init__.py: from .presets import (…)
    → presets.py: from olmo_eval.runners.asynq.batching import BatchConfig
      → runners/io/storage.py: from olmo_eval.cli.utils import console
        → triggers cli/__init__.py
          → cli/__init__.py: from olmo_eval.harness.presets import …
            ← harness.presets is half-loaded → ImportError

The fix is to relocate console to a leaf module that imports nothing from cli/harness/runners. cli.utils re-exports it so existing CLI imports are unchanged.

Files

• src/olmo_eval/common/console.py — new file, just console = Console(force_terminal=True, width=120).
• src/olmo_eval/cli/utils.py — re-exports from common.console; CLI callers unchanged.
• src/olmo_eval/runners/io/storage.py — switches to common.console. This is the actual cycle break.
• src/olmo_eval/runners/common/mixins.py — same swap.
• src/olmo_eval/common/inspection.py — replaces 7 lazy from olmo_eval.cli.utils import console as shared_console lines inside functions with one top-level import (keeps imports at top of file).

Why standalone

Split out from #159 for independent review. The fix applies cleanly to main and is a no-op for any current import path; it's preemptive infrastructure for #159.

Verification

• uv run pytest tests/ -q — 1336 passed, 22 skipped.
• make lint — clean.

[undfined]

looks good, thanks!