add hallucination metric to model evals on FMBench

[madhurprash]

FMBench can now evaluate models using a panel of LLM judges and give accuracy scores as to which candidate model is the most accurate. This issue is to do as follows:

1. Calculate a hallucination metric (that measures the amount of times a given response was actually a hallucination and incorrect)
2. Calculate the correctly "incorrect" answers, a.k.a number of times a candidate model said "i don't know" to a question rather than hallucinating and giving a response.