fuzz: keep the fuzz corpus from one scheduled run to the next #862

What is wrong

Each Fuzz run starts from the committed seeds in fuzz/seeds and discards the corpus it builds. @robinnsc found on #798 that five of the six targets were still finding new coverage in the last second of the 60-second pull request run (comment). The 600-second daily run starts from the same seeds each day, so most of each run covers the same inputs as the run before.

Where

• .github/workflows/fuzz.yml
• fuzz/seeds/

What to change

Two options:

1. An Actions cache of the corpus that only the scheduled job writes, and that pull request runs do not restore. The Fuzz workflow uses no cache, and the release workflows leave the Actions cache out on purpose because of the CodeQL cache-poisoning query. A pull request workflow can restore a cache that a run on main wrote, so any pull request could read the cached corpus.
2. A corpus that a maintainer minimizes with cargo fuzz cmin and commits to fuzz/seeds at intervals, in a pull request that gets the same review as any other change.

Recommendation

Option 2. The workflow stays free of caches, and a reviewer sees each change to the corpus.