Unstrictify expr in the bounds inference in a single pass

[vksnk]

This is to address #8686.

There is a reproducer in a bug report, but I'm not sure what would be a good way to add it? just a correctness test and if it hangs it hangs?

[abadams]

It bothers me that I don't understand why the existing behavior is exponential. I'm going to investigate more today

[abadams]

Ok I'm glad I investigated. This is not just a strict float problem - it's a problem with all pure intrinsics that get lowered in bounds inference, e.g. widening_add. The issue is that this line evaluates the bounds of the args: https://github.com/halide/Halide/blob/main/src/Bounds.cpp#L1211

But then later the intrinsic may get lowered into another form and then recursively reanalyzed, computing the bounds of the args again. If many pure intrinsics are nested, this has exponential complexity because you double the work at each call node.