Tracking issue for removing unnecessary `ty::Const::{normalize,eval}` calls from the type system

We want to remove calls to ty::Const::{normalize,eval,eval_to_*} because they won't work correctly with future reformulations of GCE, and because they're unnecessary with non-GCE consts. This issue tracks doing that so I won't forget. Boxy can probably write more motivation here idk

Boxy rationale:

with min_generic_const_args and associated_const_equality featuers normalization of type system constants will be behaving much more like types. They'll return nested goals, access in scope whjere clauses such as T: Trait<ASSOC = 10> in order to normalize instead of simply "evaluating".

This means that the only correct way to normalize a ty::Const will be to use the "normal" normalization routines such as normalize_erasing_regions or infcx/ocx/fcx.normalize. With that in mind all of the eval_x and normalize methods on ty::Const become massive footguns as they are never correct to use.

• Remove eval calls from codegen (#130644)
  • Replace calls to eval_bits with try_to_bits, which I forgot in codegen lol
• Normalize writeback results (#130645)
  • Replace calls to eval from mir_build and pattern_analysis in now that writeback results are normalized (#130715)
• Add structurally_resolve_const for use in FnCtxt/hir_typeck (#130714)
  • Replace calls to eval_* with structurally_resolve + try_to_* in hir typeck
• Audit remaining calls to eva;/normalize post-borrowck, replacing them with their try_to_* variants + possibly normalizing depending on the source of the const.

More:

• #130712

cc @BoxyUwU