PyTorch: [CI] Restrict the LF fleet to the allowlisted runner types (#200862)

Executive Summary

Impact: CI only Risk: low What Flips lf_allowlist.mode in .github/arc.yaml from all to restricted . With that, only the runner types in the allowlist can keep the lf- prefix. Any other test-matrix job that would have gone to LF gets rewritten to mt- and runs on the Meta clusters. Why Part of pytorch/ci-infra#1081 . The wiring that enforces the allowlist is already on main. This is the switch that turns it on, so day-to-day LF usage is limited to the 9 ARC targets listed in the file (the x86 avx512/amx CPU shapes, the L4/A10G GPU shapes and the default arm64 build runner) and everything else moves to Meta.

Open Silicon & Gate-Level Hardware Inspection

From a silicon and hardware engineering viewpoint: - **Interface Neutrality:** Mitigates dependency on proprietary BSPs, empowering open toolchains to address hardware directly. - **Supply Chain Verification:** Schematics and open pinouts allow deterministic audits against silicon errata. - **Lifecycle Decoupling:** Decouples physical hardware utility from commercial vendor end-of-life obsolescence.

Impact on the Open Ecosystem

Ensures that computational autonomy remains in the hands of the architect rather than closed silicon vendors.