Use CPU-only PyTorch index to speed up dev installation

[adamtheturtle]

Summary

• vws-python-mock (a dev dependency) transitively requires torch and torchvision
• By default, uv resolves these from PyPI which provides CUDA-enabled wheels on Linux (~916 MB for torch alone)
• This PR follows the faster installation guide to redirect torch and torchvision to the CPU-only PyTorch index

Does it actually make installation faster?

Yes — tested empirically (in VWS-Python/vws-python#2857):

Package	CUDA wheel (PyPI)	CPU-only wheel	Saving
torch (Linux, cp313)	916 MB	189 MB	727 MB
torchvision (Linux, cp313)	8 MB	2 MB	6 MB

~733 MB total reduction per cold-cache install on Linux CI runners.

Implementation notes

[tool.uv.sources] only applies to direct dependencies (not transitive ones), so torch and torchvision must be listed explicitly in optional-dependencies.dev for the source override to take effect. A deptry ignore is added since these packages are not directly imported in the vws-cli source code.

---

Note

Low Risk
Dependency/lockfile-only changes; main risk is unexpected dev environment resolution differences across platforms due to new index and platform markers.

Overview
Speeds up dev installs by explicitly adding torch and torchvision to optional-dependencies.dev and configuring uv ([tool.uv]) to source them from the PyTorch CPU-only wheel index instead of PyPI.

Updates dependency tooling to match this setup: adds deptry ignores for the now-explicit-but-not-imported deps, refreshes uv.lock to use platform-specific CPU wheels (dropping CUDA/NVIDIA-related packages), and bumps pyrefly plus vws-python-mock to newer versions.

^{Written by Cursor Bugbot for commit 8417eae. This will update automatically on new commits. Configure here.}