CoroSplit incorrectly assumed poison after splitting retcon coroutines with zero suspensions

As of `cefd20c496`, `CoroSplit` would incorrectly assert that the detached frame buffer to be poisoned for `retcon` coroutines with no suspensions.

Currently I am using this IR.
```llvm
target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-gnu"
declare token @llvm.coro.id.retcon(i32, i32, ptr, ptr, ptr, ptr)
declare ptr @llvm.coro.begin(token, ptr)
declare void @llvm.coro.end(ptr, i1, token)
declare noalias ptr @malloc(i64)
declare void @free(ptr)
define ptr @prototype(ptr dereferenceable(8) %buf, i1 %unwind) {
ret ptr null
}
define ptr @f(ptr dereferenceable(8) %buf) {
entry:
; We can set size to 0 and align to 1 because this coroutine has 0 suspends,
; meaning it will always be elided and requires no frame storage.
%id = call token @llvm.coro.id.retcon(i32 0, i32 1, ptr %buf, ptr @prototype, ptr @malloc, ptr @free)
%hdl = call ptr @llvm.coro.begin(token %id, ptr %buf)

; Access the frame (buffer) via handle.
; %hdl is equivalent to %buf (dereferenceable(8)).
; GEP offset is 1 (4 bytes for i32), so we access bytes 4-7, which is within the 8-byte buffer.
%val_ptr = getelementptr inbounds i32, ptr %hdl, i64 1
store i32 42, ptr %val_ptr, align 4

call void @llvm.coro.end(ptr %hdl, i1 false, token none)
ret ptr null
}
```

I ran `CoroEarly` and `CoroSplit` on it and I spotted a problem.

```sh
opt -passes='coro-early,coro-split' -S repro.ll -o repro_split.ll
```

The problem is with the `poison`

```llvm
define ptr @f(ptr dereferenceable(8) %buf) {
entry:
; We can set size to 0 and align to 1 because this coroutine has 0 suspends,
; meaning it will always be elided and requires no frame storage.
%id = call token @llvm.coro.id.retcon(i32 0, i32 1, ptr %buf, ptr @prototype, ptr @malloc, ptr @free)
br label %AllocaSpillBB

AllocaSpillBB: ; preds = %entry
br label %PostSpill

PostSpill: ; preds = %AllocaSpillBB
; BUG: %hdl replaced by poison instead of %buf,
; but %buf should have been used because it has enough frame storage and
; it is dereferenceable for 8 bytes.
%val_ptr = getelementptr inbounds i32, ptr poison, i64 1
store i32 42, ptr %val_ptr, align 4
br label %CoroEnd

CoroEnd: ; preds = %PostSpill
ret ptr null
}
```

Reproduce the issue with the provided repro.ll using opt -passes='coro-early,coro-split' -S repro.ll -o repro_split.ll, then inspect the CoroSplit transformation around the generated AllocaSpillBB and PostSpill blocks. Done means the zero-suspension retcon coroutine uses %buf rather than poison when the buffer is sufficiently dereferenceable, while preserving the existing transformation behavior.

Written by the indexing model from the issue text.