Suggest use of @helion.kernel(index_dtype=torch.int64) if index offset is out of bound for int32

[yf225]

Fixes #686.

[jansel]

We also have @helion.kernel(index_dtype=torch.int64). It might be better to error and force people to provide that annotation. Otherwise, you might get a different kernel based on the size hints of the inputs, and I don't think we will automatically recompile if you run first with a small size then a large one.