Retry transient Zig ARM64 release allocation failures

[ajroetker]

Summary

• wrap Zig release build steps so their first-attempt output can be classified
• retry once only for aarch64-linux-musl + ReleaseSmall allocation failures
• preserve the populated local cache for the retry
• fail immediately for every unrelated compiler or source error

Why

The ARM64 release compile repeatedly aborts in Zig/LLVM allocation handling despite zero kernel/cgroup OOM events. A 64 GB runner failed too. In the isolated diagnostic, the fresh release compile failed with LLVM ERROR: out of memory / Buffer allocation failed, then both the full build-runner replay and direct compiler replay succeeded from independent copies of the partially populated local cache at essentially the same ~15.9 GB max RSS and zero swap.

Diagnostic evidence: #450 and cache A/B run.

This is a bounded release mitigation, not a claim that the upstream Zig/LLVM issue is fixed.

Verification

• bash -n scripts/packaging/build_zig_release_archive.sh
• git diff --check
• fake-compiler wrapper test: recognized allocation failure called install twice, then lite-capi, and produced a valid archive
• fake-compiler wrapper test: unrelated compilation failure called install once and exited 1 without retrying

[ajroetker]

End-to-end ARM64 validation passed on the original normal runner: run 31059377379.

Observed real wrapper behavior:

1. First install failed with std::bad_alloc.
2. The new classifier matched it and emitted the bounded retry.
3. Second install succeeded using the populated local cache.
4. lite-capi succeeded without a retry.
5. Staging, archive content validation, and artifact upload all succeeded.

The runner remained the original 32 GB machine / 24,000 Mi request. PR checks are green and the PR is mergeable.