Suboptimal diagnostic if const eval error site is in a different crate

[oli-obk]

We currently produce fun diagnostics like

error: any use of this value will cause an error
  --> $SRC_DIR/core/src/intrinsics.rs:LL:COL
   |
LL |     unsafe { copy_nonoverlapping(src, dst, count) }
   |              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
   |              |
   |              memory access failed: pointer must be in-bounds at offset 8, but is outside bounds of alloc6 which has size 4
   |              inside `copy_nonoverlapping::<u32>` at $SRC_DIR/core/src/intrinsics.rs:LL:COL
   |              inside `std::ptr::read::<u32>` at $SRC_DIR/core/src/ptr/mod.rs:LL:COL
   |              inside `ptr::mut_ptr::<impl *mut u32>::read` at $SRC_DIR/core/src/ptr/mut_ptr.rs:LL:COL
   |              inside `_MUT_READ` at $DIR/out_of_bounds_read.rs:15:37
   | 
  ::: $DIR/out_of_bounds_read.rs:15:5
   |
LL |     const _MUT_READ: u32 = unsafe { (PAST_END_PTR as *mut u32).read() };
   |     --------------------------------------------------------------------

Which I have multiple problems with (all of which are my fault 😛)

1. the main message only makes sense when you realize this is a lint
2. the diagnostic points to another crate's internals as the main span
3. the local crate's span is pointing at the entire item, instead of the useful span that we have (last element in the stack trace)

I think we should adjust the diagnostic to report on the site in the currently compiling crate while still mentioning the backtrace in full, so something like

error: memory access failed: pointer must be in-bounds at offset 8, but is outside bounds of alloc6 which has size 
  --> $SRC_DIR/core/src/intrinsics.rs:LL:COL
   |
LL |     const _MUT_READ: u32 = unsafe { (PAST_END_PTR as *mut u32).read() };
   |                                                               ^^^^^^^ in function call here
   |     
LL |     unsafe { copy_nonoverlapping(src, dst, count) }
   |              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ failure in external crate here
   |              inside `copy_nonoverlapping::<u32>` at $SRC_DIR/core/src/intrinsics.rs:LL:COL
   |              inside `std::ptr::read::<u32>` at $SRC_DIR/core/src/ptr/mod.rs:LL:COL
   |              inside `ptr::mut_ptr::<impl *mut u32>::read` at $SRC_DIR/core/src/ptr/mut_ptr.rs:LL:COL
   |              inside `_MUT_READ` at $DIR/out_of_bounds_read.rs:15:37

cc @rust-lang/wg-diagnostics @rust-lang/wg-const-eval

[RalfJung]

Also see #71800: the lint should hopefully become a hard error eventually. I am currently working on a PR to make it a future-incompat lint, as a first step.

[RalfJung]

Some other notes:

• Following [const_panic] Report const_panic diagnostics identically to compiler_error invocations #79695, the main message should probably be something more specific than what it currently is
• Even within the same crate, it can be confusing when the main span is not within the body of the offending const, but in some const fn who-knows-where.