Reconsider caching the full Swift toolchain (~1 GB per version) + document matrix cost #113

Problem

Consumer repos each carry ~1 GB Swift toolchain caches per Swift version (e.g. swift-6.3-ubuntu2404 ≈ 1015 MB, swift-6.2.3-ubuntu2404 ≈ 957 MB). Repos testing two Swift versions pay ~2 GB just for toolchains, and these are saved on PR refs too (see #111), so they multiply.

Questions / proposal

• Is caching the full toolchain tarball worth it vs. pulling from swift.org / using the official Swift container image, which is often comparably fast and costs zero cache quota?
• If we keep toolchain caching, it should only be saved on the default branch (covered by Gate cache saves to the default branch (eliminate PR-stranded caches) #111) and ideally shared, not re-saved per PR.
• Consider documenting the matrix cost: each (OS × Swift version × Windows-server image) combination multiplies cache count. Trimming Swift versions / Windows images the project doesn't need to support is a quick win.

Acceptance criteria

• Decision recorded on whether to keep toolchain caching by default
• If kept, toolchain caches are not written on PR refs
• README notes the matrix-multiplication storage effect with guidance to trim unused combinations