chore: clarify platform-specific torch installation defaults

[mxj711]

Summary

This PR standardizes UniLab's default PyTorch installation behavior across platforms and updates the related developer-facing documentation.

Changes

• Simplify Linux installation from multiple CUDA extras to a single default uv sync flow
• Make Linux default resolve to PyTorch cu128 wheels through tool.uv.sources
• Keep macOS on the default PyPI torch wheels
• Add torchvision as a direct dependency and align lockfile resolution with the new source setup
• Remove obsolete CUDA optional extras (cu118, cu124, cu126, cu128) from project metadata
• Update README, CONTRIBUTING, and Getting Started docs in English, Chinese, Japanese, and Korean
• Narrow make test / make test-cov / make test-fast to exclude veryslow cases explicitly
• Trim outdated Go2 reward fixture fields in tests/conftest.py

Why

Previously, Linux installation instructions required users to choose a CUDA extra manually, while the actual dependency setup was moving toward a more opinionated default. This change makes the installation path clearer and reduces setup ambiguity:

• macOS users continue to install directly from PyPI
• Linux users get the default cu128 PyTorch wheel set without extra flags
• docs and lockfile now reflect the real dependency behavior

Impact

• Developer setup becomes simpler on Linux
• Existing commands using uv sync --extra cu124-style flows are no longer the documented or supported default
• CI / local test commands now distinguish regular non-slow tests from veryslow smoke coverage more explicitly

Validation

• Updated dependency metadata in pyproject.toml
• Regenerated uv.lock
• Synced installation docs across all supported languages