Extend fuzzer to mutate existing examples

Currently the fuzzer just builds examples from scratch. This results in a program distribution that is different than might appear in human written programs generally with less structure. For a more representative distribution, the fuzzer could also take existing examples and mutate them.

Possible interesting mutations:
* replace random uses in the graph with literals or new parameters
* Randomly rewire the graph by redirecting a use to a new source of the same type. Have to be careful that cycles aren't introduced.
* Randomly change opcodes with compatible types and operand types.