compat: add non-determinism into negative compat checks

[dotnwat]

This is super cool but I think there should some sort of non determinism built into this..meaning we should choose to skip perturbing the input with some probability. I think so because the default equality comparison short circuits the remaining checks at first mismatch.. Example.

class Foo { int a, complex_type b. }

If we always perturb a, we don't find any bugs in b? (or am I totally missing something here?).

Originally posted by @bharathv in https://github.com/redpanda-data/redpanda/pull/5880#discussion_r939821134