This is a much simpler way to achieve the pre-#86003 behavior of mem::replace not needing dynamically-sized memcpys (at least before inlining), than re-doing #81238 (which needs #86699 or something similar).

I didn't notice it until recently, but ptr::write already explicitly avoided using the wrapper, while ptr::read just called the wrapper (and was the reason for us observing any behavior change from #86003 in Rust-GPU).

The codegen test I've added fails without the change to core::ptr::read like this (ignore the v0 mangling, I was using a worktree with it turned on by default, for this):

       13: ; core::intrinsics::copy_nonoverlapping::<u8>
       14: ; Function Attrs: inlinehint nonlazybind uwtable
       15: define internal void @_RINvNtCscK5tvALCJol_4core10intrinsics19copy_nonoverlappinghECsaS4X3EinRE8_25mem_replace_direct_memcpy(i8* %src, i8* %dst, i64 %count) unnamed_addr #0 {
       16: start:
       17:  %0 = mul i64 %count, 1
       18:  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 1 %dst, i8* align 1 %src, i64 %0, i1 false)
not:17      !~~~~~~~~~~~~~~~~~~~~~                                                                     error: no match expected
       19:  ret void
       20: }

With the core::ptr::read change, core::intrinsics::copy_nonoverlapping doesn't get instantiated and the test passes.

r? @m-ou-se cc @nagisa (codegen test) @oli-obk / @RalfJung (miri diagnostic changes)