[Feature] Support more inference "corner cases"

[mohammadfawaz]

#28642 and all the work that came before support a good amount of common cases. We can do better thought. Here are some example that I would like to support in the future:

1. Implicit array indices. Maybe default to u32 since I think we cast to u32 anyways?
2. In let x = 1u32 + (true ? 1 : 2), we should infer 1 and 2 to be u32.
3. In let x = [1, 2u32, 3], we should infer 1, and 3 to be u32.
4. In let x = 1; let y = x + 3u32, we should infer that x and 1 are both u32. This requires revisiting previously type checked statements.
5. In let x = 1 + 1 + 1u32, we should infer that the 1s are both u32. The problem here is that 1 + 1 is visited first and we quickly give up.
6. In for i in 0u32..4 { }, we should infer that the 4 is a u32
7. In assert_eq(1u32, 1), we should infer that the 1 is a u32.
8. Mapping function such as .set(0, 0) should work.
9. More to come

I have some ideas around these but they require a bigger refactor of parts in the compiler. I think we need a way to visit a statement/expression and have the ability to visit it again when we gather more information. So, certainly a way to throw away collected errors and not always emit them in case we revisit an item again.