fraud: Bad encoding fraud proofs for unordered share

[rootulp]

Context

While reviewing the NMT, @evan-forbes raised the following scenario:

A malicious 2/3 majority of consensus nodes confirm a block that has a data square with one share that isn't lexicographically ordered. It is possible for light nodes to miss sampling the one unordered share. Since light nodes don't reconstruct the entire data square, they need to be made aware of such behavior (presumably via a bad encoding fraud proof).

Problem

It isn't clear that the current implementation of bad encoding fraud proofs work for this scenario.

Proposal

1. Update NMT to handle this scenario. See Defining and Clarifying the Verification Process for NMT Proofs Generated from Non-Lexicographically Ordered Trees nmt#97
2. Investigate the changes needed (likely in this repo) to support a bad encoding fraud proof for a data square with one unordered share
3. Improve unit test coverage to account for this scenario (related: eds/byzantine: Bad encoding fraud proofs need more unit testing #1086)