[CoroEarly][IR] Clarify semantic of llvm.coro.end (#191752)

We introduced a workaround for the following pattern in #139243:
``` LLVM
define void @fn() presplitcoroutine {
  %__promise = alloca ptr, align 8
  ...

coro.ret:
  call void @llvm.coro.end(ptr null, i1 false, token none)
  store ptr null, ptr %__promise, align 8
  ret void
}
```
where DSE considers `__promise` dead after the return and eliminates the
store, leading to a miscompilation.

However, after #151067, the problematic pattern is gone. And it
currently looks like:
``` LLVM
gro.conv:
  store ptr null, ptr %__promise, align 8
  br label %after.gro.conv

after.gro.conv:
  br i1 %body.done, label %coro.cleanup, label %coro.ret

coro.ret:
  call void @llvm.coro.end(ptr null, i1 false, token none)
  ret void
```
DSE cannot eliminate the store because `__promise` escapes through
`coro.id`, and `coro.end` post-dominates the store. It turns out that
accessing the coroutine frame after `coro.end` is not safe. This patch
proposes clarifying the semantics of `coro.end` and reverting the
workaround, as it confuses alias analyses. I have scanned the tests and
updated the uses of the coroutine frame after `coro.end`.