ci: speed up E2E pipeline with single-arch builds, registry cache, and skip redundant cluster rebuild

[drew]

Summary

Three fixes to address ~10 minute E2E pipeline regression:

• Single-arch E2E builds: E2E workflow now builds amd64-only images instead of multi-arch (linux/amd64,linux/arm64). The arm64 cross-compilation via QEMU was adding ~7 min per build job. Publish workflow retains multi-arch via the default.
• Restore GHCR registry build cache: Re-adds DOCKER_CACHE_FROM/DOCKER_CACHE_TO registry cache that was removed in chore: remove unnecessary cache config #54. Without it, every CI Docker build compiles from scratch since the local filesystem cache is skipped when CI=true.
• Skip redundant cluster image rebuild: The E2E bootstrap step was rebuilding the cluster image locally (including a full Rust compilation of navigator-sandbox ~2 min) even though it was already pre-built and pulled from GHCR. Sets SKIP_CLUSTER_IMAGE_BUILD=1.

Expected Impact

Fix	Time Saved
Single-arch E2E builds	~5-7 min
GHCR registry cache	~2-5 min
Skip cluster rebuild	~2.5 min
Total	~10-12 min

Test Plan

• PR has e2e label to trigger E2E pipeline — verify it passes and is faster than recent runs (~22 min -> ~10-12 min target)
• Publish workflow is unaffected (uses default multi-arch platform)