[Feat] Add text-board Sokoban agent with Valen levels

[hsliuustc0106]

Adds s1a run sokoban so the existing decision-model loop can play Sokoban with reproducible levels and ordinary episode reports. It adapts Valen's standard-library simulator and bundles its 100 selected evaluation levels; the model receives an ASCII board, the remaining step budget and four directional choices. A solved game scores 1, otherwise 0.

Blocked moves consume steps, the environment caps episodes at 200 actions, and level selection uses --seed as an offset without wrapping. Rethink and rule mode are rejected: the agent receives no solver assistance or reference solutions. This is a text-board adaptation, not a reproduction of Valen's visual benchmark, and the supplied levels are a selected subset. No learned-model accuracy or performance claim is made.

Code and data revisions, Apache-2.0 attribution, CLI usage and evaluation limits are documented. Evaluator-only reference solutions are confined to test fixtures.

Validation:

• Full pytest suite: 502 passed, 36 skipped (optional dependencies/platform checks), 80 subtests passed.
• All 100 upstream reference solutions replay to completion; transition, reset, step-limit, selection and offline DeepAgent integration tests pass.
• Ruff formatting/lint, ty with the reused development environment, CLI/MCP smoke checks and uv build pass.
• A 10-step random-policy CLI run completed with zero errors, invalid keys or model calls; no GPU execution.

Try from the checkout:

uv run s1a run sokoban --model random --rethink off --episodes 1 --max-steps 10
uv run s1a run sokoban --model jev --rethink off --episodes 100