feat: add optional CPU-only PyTorch extra and runtime guard

[fmueller]

Motivation

• Make PyTorch an optional dependency so translation features are opt-in and the base install avoids heavy GPU builds.
• Provide a pinned, CPU-only PyTorch extra for Linux x86_64 to allow users to avoid CUDA downloads when they only need CPU inference.
• Surface a clear runtime error when PyTorch is missing that explains how to install the translation extras.

Description

• Move torch out of core dependencies into optional extras translation and translation-cpu in pyproject.toml with pinned CPU wheel URLs for targeted Linux/Python combos and a fallback spec for other platforms.
• Add a _require_torch helper in src/scribae/translate/mt.py that checks for torch via importlib.util.find_spec and raises a user-facing RuntimeError with installation guidance if missing, and use the returned module for device detection.
• Update translation pipeline code to lazily require torch once a translation pipeline is needed and avoid importing it at CLI startup.
• Document the new install flow in README.md and update the Quick start to show uv sync --locked --dev --extra translation and --extra translation-cpu usage.

Testing

• Ran uv run ruff check --fix, which failed due to inability to download the pinned CPU torch wheel from download.pytorch.org during environment sync.
• Ran uv run mypy, which failed for the same reason related to dependency fetch errors for the pinned torch wheel.
• Ran uv run pytest, which also failed because the environment sync attempted to fetch the pinned CPU torch wheel and the download failed.

---

Codex Task