LLVM does not know that the multiplication never overflows, which causes
it to generate unnecessary instructions. Use `usize::unchecked_mul`, so
that it can fold the `dst_cap` calculation when `size_of::<I::SRC>() ==
size_of::<T>()`.

Running:

```
rustc -C llvm-args=-x86-asm-syntax=intel -O src/lib.rs --emit asm`
```

```rust

pub struct Foo([usize; 3]);

pub fn unwrap_copy(v: Vec<Foo>) -> Vec<[usize; 3]> {
    v.into_iter().map(|f| f.0).collect()
}
```

Before this commit:

```
define void @unwrap_copy(ptr noalias nocapture noundef writeonly sret([24 x i8]) align 8 dereferenceable(24) %_0, ptr noalias nocapture noundef readonly align 8 dereferenceable(24) %iter) {
start:
  %me.sroa.0.0.copyload.i = load i64, ptr %iter, align 8
  %me.sroa.4.0.self.sroa_idx.i = getelementptr inbounds i8, ptr %iter, i64 8
  %me.sroa.4.0.copyload.i = load ptr, ptr %me.sroa.4.0.self.sroa_idx.i, align 8
  %me.sroa.5.0.self.sroa_idx.i = getelementptr inbounds i8, ptr %iter, i64 16
  %me.sroa.5.0.copyload.i = load i64, ptr %me.sroa.5.0.self.sroa_idx.i, align 8
  %_19.i.idx = mul nsw i64 %me.sroa.5.0.copyload.i, 24
  %0 = udiv i64 %_19.i.idx, 24
  %_16.i.i = mul i64 %me.sroa.0.0.copyload.i, 24
  %dst_cap.i.i = udiv i64 %_16.i.i, 24
  store i64 %dst_cap.i.i, ptr %_0, align 8
  %1 = getelementptr inbounds i8, ptr %_0, i64 8
  store ptr %me.sroa.4.0.copyload.i, ptr %1, align 8
  %2 = getelementptr inbounds i8, ptr %_0, i64 16
  store i64 %0, ptr %2, align 8
  ret void
}
```

After:

```
define void @unwrap_copy(ptr noalias nocapture noundef writeonly sret([24 x i8]) align 8 dereferenceable(24) %_0, ptr noalias nocapture noundef readonly align 8 dereferenceable(24) %iter) {
start:
  %me.sroa.0.0.copyload.i = load i64, ptr %iter, align 8
  %me.sroa.4.0.self.sroa_idx.i = getelementptr inbounds i8, ptr %iter, i64 8
  %me.sroa.4.0.copyload.i = load ptr, ptr %me.sroa.4.0.self.sroa_idx.i, align 8
  %me.sroa.5.0.self.sroa_idx.i = getelementptr inbounds i8, ptr %iter, i64 16
  %me.sroa.5.0.copyload.i = load i64, ptr %me.sroa.5.0.self.sroa_idx.i, align 8
  %_19.i.idx = mul nsw i64 %me.sroa.5.0.copyload.i, 24
  %0 = udiv i64 %_19.i.idx, 24
  store i64 %me.sroa.0.0.copyload.i, ptr %_0, align 8
  %1 = getelementptr inbounds i8, ptr %_0, i64 8
  store ptr %me.sroa.4.0.copyload.i, ptr %1, align 8
  %2 = getelementptr inbounds i8, ptr %_0, i64 16
  store i64 %0, ptr %2, align 8, !alias.scope !9, !noalias !14
  ret void
}
```

Note that there is still one more `mul,udiv` pair that I couldn't get
rid of. The root cause is the same issue as rust-lang#121239, the `nuw` gets
stripped off of `ptr::sub_ptr`.

`Iterator::try_fold` gets called on the underlying Iterator in
`SpecInPlaceCollect::collect_in_place` whenever it does not implement
`TrustedRandomAccess`. For types that impl `Drop`, LLVM currently can't
tell that the drop can never occur, when using the default
`Iterator::try_fold` implementation.

For example, the asm from the `unwrap_clone` method is currently:

```
unwrap_clone:
        push    rbp
        push    r15
        push    r14
        push    r13
        push    r12
        push    rbx
        push    rax
        mov     rbx, rdi
        mov     r12, qword ptr [rsi]
        mov     rdi, qword ptr [rsi + 8]
        mov     rax, qword ptr [rsi + 16]
        movabs  rsi, -6148914691236517205
        mov     r14, r12
        test    rax, rax
        je      .LBB0_10
        lea     rcx, [rax + 2*rax]
        lea     r14, [r12 + 8*rcx]
        shl     rax, 3
        lea     rax, [rax + 2*rax]
        xor     ecx, ecx
.LBB0_2:
        cmp     qword ptr [r12 + rcx], 0
        je      .LBB0_4
        add     rcx, 24
        cmp     rax, rcx
        jne     .LBB0_2
        jmp     .LBB0_10
.LBB0_4:
        lea     rdx, [rax - 24]
        lea     r14, [r12 + rcx]
        cmp     rdx, rcx
        je      .LBB0_10
        mov     qword ptr [rsp], rdi
        sub     rax, rcx
        add     rax, -24
        mul     rsi
        mov     r15, rdx
        lea     rbp, [r12 + rcx]
        add     rbp, 32
        shr     r15, 4
        mov     r13, qword ptr [rip + __rust_dealloc@GOTPCREL]
        jmp     .LBB0_6
.LBB0_8:
        add     rbp, 24
        dec     r15
        je      .LBB0_9
.LBB0_6:
        mov     rsi, qword ptr [rbp]
        test    rsi, rsi
        je      .LBB0_8
        mov     rdi, qword ptr [rbp - 8]
        mov     edx, 1
        call    r13
        jmp     .LBB0_8
.LBB0_9:
        mov     rdi, qword ptr [rsp]
        movabs  rsi, -6148914691236517205
.LBB0_10:
        sub     r14, r12
        mov     rax, r14
        mul     rsi
        shr     rdx, 4
        mov     qword ptr [rbx], r12
        mov     qword ptr [rbx + 8], rdi
        mov     qword ptr [rbx + 16], rdx
        mov     rax, rbx
        add     rsp, 8
        pop     rbx
        pop     r12
        pop     r13
        pop     r14
        pop     r15
        pop     rbp
        ret
```

After this PR:

```
unwrap_clone:
	mov	rax, rdi
	movups	xmm0, xmmword ptr [rsi]
	mov	rcx, qword ptr [rsi + 16]
	movups	xmmword ptr [rdi], xmm0
	mov	qword ptr [rdi + 16], rcx
	ret
```

Fixes rust-lang#120493

LLVM currently adds a redundant check for the returned option, in addition
to the `self.ptr != self.end` check when using the default
`Iterator::fold` method that calls `vec::IntoIter::next` in a loop.