[Deps] Require laya 0.3.9 and drop the local weight-init skip

[Yunaik]

Why

#22 made --model laya load in about 3 s by running laya.load under transformers' no_init_weights (without_weight_init() in s1a/decision_models/laya.py). Laya 0.3.9, published 2026-09-23, does the same inside laya.load: it builds the model under no_init_weights with the same transformers.initialization / transformers.modeling_utils fallback, checks the checkpoint with _verify_compatibility, then loads it with strict=True (laya/agent.py:327-335 in 0.3.20).

uv.lock pins laya 0.3.5, which predates that change. The before/after numbers in #22 were measured on 0.3.5.

What

1. Set the extra to laya = ["laya>=0.3.9"] in pyproject.toml and run uv lock --upgrade-package laya.
2. Remove without_weight_init() and the with block around laya.load in LayaModel.from_env.
3. Remove the transformers.initialization stubs: TestFromEnv.setUp and the three helper tests in tests/test_decision_models_laya.py, and the stub in tests/test_decision_models_factory.py.
4. Reword the [Fix] Load Laya without the throwaway random weight init #22 entry in CHANGELOG.md to name the version bump.

After this, laya checks that the checkpoint fills every weight. The tests no longer stub transformers.

Verification

I have not timed a load on laya 0.3.9 or later.

• Repeat the [Fix] Load Laya without the throwaway random weight init #22 measurement on the bumped version: a cold LayaModel.from_env() in a fresh process on CPU with transformers 5.17.0, and s1a decide --model laya end to end. The expected load is about 3 s.
• The upgrade includes every laya change from 0.3.6 to the locked version (for example Agent.accelerate()). Compare answers on the default checkpoint against 0.3.5.
• uv run pytest -q on the core install and with uv sync --extra laya.