The problem was that in the by-address emission, we were calling
`getAddressForInPlaceInitialization` twice, triggering the assert.
The first time in `emitExprInto` for the normal result case.
The second time to obtain the address again when generating the
catch block to inject a `.none` into that same address.
This patch does a light refactoring to more closely mirror
`visitOptionalEvaluationExpr`, which avoids calling the asserting method.
fixes rdar://80277465
