Note [Patching magic definitions]

We sometimes need to have access to defined Ids in pure contexts. Usually, we
simply "wire in" these entities, as we do for types in GHC.Builtin.Types and for Ids
in GHC.Types.Id.Make. See Note [Wired-in Ids] in GHC.Types.Id.Make.

However, it is sometimes *much* easier to define entities in Haskell,
even if we need pure access; note that wiring-in an Id requires all
entities used in its definition *also* to be wired in, transitively
and recursively.  This can be a huge pain.  The little trick
documented here allows us to have the best of both worlds.

Motivating example: unsafeCoerce#. See [Wiring in unsafeCoerce#] for the
details.

The trick is to

* Define the known-key Id in a library module, with a stub definition,
     unsafeCoerce# :: ..a suitable type signature..
     unsafeCoerce# = error "urk"

* Magically over-write its RHS here in the desugarer, in
  patchMagicDefns.  This update can be done with full access to the
  DsM monad, and hence, dsLookupGlobal. We thus do not have to wire in
  all the entities used internally, a potentially big win.

  This step should not change the Name or type of the Id.

Because an Id stores its unfolding directly (as opposed to in the second
component of a (Id, CoreExpr) pair), the patchMagicDefns function returns
a new Id to use.

Here are the moving parts:

- patchMagicDefns checks whether we're in a module with magic definitions;
  if so, patch the magic definitions. If not, skip.

- patchMagicDefn just looks up in an environment to find a magic defn and
  patches it in.

- magicDefns holds the magic definitions.

- magicDefnsEnv allows for quick access to magicDefns.

- magicDefnModules, built also from magicDefns, contains the modules that
  need careful attention.