Add NaN/Inf checking mode

[akhilg-nv]

While debugging accuracy issues in models (e.g. FP16 diffusion) it is useful to be able to identify the layer/operation where error is being introduced. Currently this has to be done manually while running in lazy mode, which can require non-trivial code changes and significantly slows down evaluation. Ideally we should have a solution that makes this possible with minimal code change and even in compiled mode.