ci: Dogfood Cargo target restoration (#13336)
## Why
A six-run benchmark showed distributed sccache was 13.4% slower than
clean compilation, while restoring a full Cargo target was 70.3% faster
including download and rebuild. Real workflow dogfooding confirmed the
target strategy materially improves Rust tests.
## What
Disable repository sccache dogfooding and restore/save full Ubuntu Cargo
`target/` state around Rust tests.
## How
- Only trusted `main` pushes publish target snapshots; PRs are
restore-only and forks cannot write main-scoped state.
- Keys segment OS, architecture, toolchain/configuration, lockfile, and
Rust workspace inputs; compatible prior source states are fallback
restores.
- Exact cache-key hits skip redundant uploads. A total main miss forces
the Rust task once so the snapshot is populated even if Turbo has an
exact task artifact.
- `Swatinem/rust-cache` is disabled for the Rust test matrix so
workspace and incremental target artifacts are preserved.
- Turbo Remote task caching remains unchanged.
Measured Ubuntu source-change dogfood:
- Target restore: 15s for 931MB
- Cargo compile: 2m25s → 48s
- Rust task: 3m32s → 1m59s
- Full job: 4m54s → 3m27s
- Updated snapshot save: 8s