Build corpus images last, in a separate workflow step

[tadamcz]

The agent_corpus target is exceptionally slow (~9 GB corpus fetch + corpus build). This splits the build loop in two:

• Build and push images (no corpus) — agent and comparator for every FC pin
• Build and push corpus images — agent_corpus for every FC pin, always last

The discovery step now emits two lists (builds / corpus_builds); the build logic is otherwise unchanged, and the Dockerfile is untouched.

Benefits:

• The fast images are pushed as early as possible instead of waiting behind the corpus per-commit.
• The Actions step list (and gh run view --job) shows at a glance when everything except the corpus is done.

Both steps run in the same job on the same buildx builder, so agent_corpus still gets its FROM agent layers as cache hits from the first step — only the corpus stages actually run in step two.