Improve warmup checking for max new tokens when using speculative decoding

[tgaddair]

If speculative decoding is in use and the user wants to generate up to the max positional embeddings of the model, errors can arise at runtime causing a CUDA device-side assert error. We should do a better job detecting these errors during warmup, or gracefully handling this edge case per request.

[AmanSCoder]

Is this issue still unresolved/valid? If it is still unresolved/valid, I would like to work on it.