[Liveness][analyzer] Fix handling of [[assume]] attributes (#198618)
Before this commit, if the analyzer encountered code like
```
int f(int a, int b) {
[[assume(a == 2), assume(b == 3)]];
return a + b;
}
```
it performed the following steps:
1. It visited the expression `a == 2` with `ExprEngine::Visit` (after
visiting its sub-expressions, within the regular visitation that visits
each statement of the `CFGBlock`). This triggered the `EagerlyAssume`
logic and separated two execution paths.
2. It discarded the result bound to `a == 2` from the `Environment`
because `a == 2` is not a direct child of the `AttributedStmt`.
3. Analogously, it visited an evaluated `b == 3`.
4. Analogously, it discarded the result bound to `b == 3`.
5. On each execution path `VisitAttributedStmt` was reached, it ran the
`PreStmt<AttributedStmt>` checkers and stored their results in
`CheckerPreStmt`. As there are AFAIK no such checkers, after this step
the only element of `CheckerPreStmt` was the node `Pred`.
6. It constructed a `NodeBuilder`, which was unused, but its constructor
inserted the nodes in `CheckerPreStmt` into `EvalSet`. This was
essentially a shortcut that skipped the handling of the assumptions --
but ultimately its effects were absorbed and negated by the tricky side
effects of `ExprEngine::Visit`.
7. For each node `N` in `CheckerPreStmt` (i.e. for `N == Pred`) and for
each assume attribute, it created a child node of `N` where the
assumption of **that one attribute** was visited again with
`ExprEngine::Visit`. (I.e. it separated execution paths and implicitly
placed "OR" relationships between the assumptions.)
8. It **completely ignored** the value produced by the assumption
expression: did not update the state and did not discard execution paths
where the [[assume]] assumption contradicted earlier knowledge.
This commit fixes all four logic errors:
1. The liveness analysis is updated to ensure that an `AttributedStmt`
keeps alive the assumption expressions of its [[assume]] attributes.
This ensured that their results remain in the `Environment`.
2. Now there is no `NodeBuilder` that skips the handling the
assumptions.
3. The invocation of `ExprEngine::Visit` is removed from
`VisitAttributedStmt`: there is no reason to revisit those statements
(now that we don't lose the evaluation results).
4. The state is updated to record that the values of the assumption
expressions are ALL assumed to be true. If the assumption leads to a
contradiction, the execution path is discarded (by not generating a
node).
This commit also adds a few new testcases and resolves some FIXME
comments in previously existing tests.