Directly call arithmetic primops instead of indirecting through builtins

[gustavderdrache]

Motivation

Right now, some (but not all) of the Nix language's arithmetic expressions can be shadowed by overriding definitions:

$ nix repl
Nix 2.28.3
Type :? for help.
nix-repl> let __sub = throw "bogus"; in 2 - 2 
error:
       … while calling the 'throw' builtin
         at «string»:1:13:
            1| let __sub = throw "bogus"; in 2 - 2
             |             ^

       error: bogus

As of this PR, this is no longer possible:

$ nix repl
Nix 2.28.3
Type :? for help.
nix-repl> let __sub = throw "bogus"; in 2 - 2
0

In addition, stack traces are annotated with "while evaluating the %s operator" for symmetry with the "while calling the '%s' builtin" from the current implementation:

Old:

nix-repl> 2 - (throw "bogus")
error:
       … while calling the 'sub' builtin
         at «string»:1:3:
            1| 2 - (throw "bogus")
             |   ^

       … while calling the 'throw' builtin
         at «string»:1:6:
            1| 2 - (throw "bogus")
             |      ^

       error: bogus

New:

nix-repl> 2 - (throw "bogus")
error:
       … while evaluating the - operator
         at «string»:1:3:
            1| 2 - (throw "bogus")
             |   ^

       … while calling the 'throw' builtin
         at «string»:1:6:
            1| 2 - (throw "bogus")
             |      ^

       error: bogus

Context

This PR adds binary operation classes for the remaining builtins. As mentioned above, evaluation exceptions thrown have traces added for parity with the old reference to builtin calls. This somewhat inconsistent with other operators: 2 + (throw "bogus") does not mention the operator context. If this PR looks good, I can do a developer experience pass to get the error messages working.

While I've factored out the arithmetic builtins, I haven't touched prim_lessThan because of how it's handled within prim_sort. I'd like to factor it out so that we can just directly use comparison operators instead of essentially backporting the old logic within the AST - maybe just use an enum to get a comparator.

I haven't updated the tests. There are a few failures due to my having changed the error messages. I would like to see if this code is good before finalizing the error messages expected by the test suite.

[piegamesde]

While we can probably all agree on the end goal and that going in this direction is desirable, it should be noted that this will cause existing code to silently evaluate differently than it did before. This may or may not lead to catastrophic circumstances on some end user's machine. For Lix, we instead chose the approach of forbidding the offensive overrides, which I'd recommend taking. This means that instead of silently changing semantics, code with operator overrides will now error out. (Feel free to take the commit, just make sure to also grab its fixup commit done a bit later which relaxes the rules for __nixPath and __findFile again)