We should add fuzzer hook for constructing arbitrary compressed array

[joseph-isaacs]

I think the fuzzer is doing a bad job exploring encodings that are unlikely to be chosen from random Canonical compression (e.g. its unlikely for the fuzzer to generate complex ALP arrays).

Add a arbitrary impl for these arrays and hook them into the array_ops path.

[robert3005]

I think it would be simpler to create a targeted fuzzer at some encodings or dtypes. In this case generating floats with some bias towards pseudodecimals is likely the way to go