[dev.fuzz] internal/fuzz: don't store corpus in memory

[rolandshoemaker]

Currently the coordinator will read all on the input corpus into memory, and store any newly discovered inputs in memory as well. For large corpora this will result in significant memory usage (see #44517).

Instead the coordinator should only read inputs from disk when necessary, storing a list of input paths that are available, and only hold them as long as is necessary to send the input to a worker.

One open question is how to handle inputs added via f.Add, since they only ever existed in memory. We could add these to the corpus cache, so there are files for them, or store just these values in memory (in CorpusEntry have either value or filepath maybe?) and document that if you have a lot of these kinds of inputs you should store them as files rather than adding them via f.Add (I've seen a number of targets which decompress archives of inputs and then add them via f.Add, so this may not be viable for everyone).

[gopherbot]

Change https://golang.org/cl/345096 mentions this issue: [dev.fuzz] internal/fuzz: don't store corpus in memory