Faster, more reliable selective test shard downloads

Selective test shard downloads are now more reliable and leaner. Each shard only downloads the test products it actually needs (shared binaries once, plus per-module test bundles for the suites assigned to it), instead of every shard pulling the full bundled artifact. The download itself now retries automatically, so the previous wall-clock timeout failure on large shard bundles no longer hard-stops a CI run. On real workloads this reduces total test artifact egress while keeping shard run-time balance identical. The pipeline also stops running out of memory when archiving and uploading very large test bundles (hundreds of modules, multi-GB), so selective testing works on big projects.