RFC: stop wrapping unreachable statements in Const

[JeffBezanson]

This does not make sense semantically, and we have better sources of reachability info now. The biggest concrete problem is that it makes code_warntype output confusing, e.g.

1 ─ %1 = Main.throw::Core.Const(throw)
│   %2 = Main.AssertionError::Core.Const(AssertionError)
│   %3 = Base.string("x1=", x1)::String
│   %4 = (%2)(%3)::Core.PartialStruct(AssertionError, Any[String])
│        (%1)(%4)
└──      Core.Const(:(return %5))

where the final unreachable return statement looks like a type.

What I have here seems to work (passes Compiler tests) but I need some help because I don't think it's optimal. The change in construct_ssa! is not quite right since iterated_dominance_frontier is now reaching unreachable statements that it skipped before, but I haven't yet figured out how it's depending on the Const wrapping for that.

[aviatesk]

The change in construct_ssa! is not quite right since iterated_dominance_frontier is now reaching unreachable statements that it skipped before, but I haven't yet figured out how it's depending on the Const wrapping for that.

I think we likely need to tweak convert_to_ircode, in particular

julia/Compiler/src/optimize.jl

Lines 1247 to 1252 in b61664e

	if !(idx < length(code) && isa(code[idx + 1], ReturnNode) && !isdefined((code[idx + 1]::ReturnNode), :val))
	# Any statements from here to the end of the block have been wrapped in Core.Const(...)
	# by type inference (effectively deleting them). Only task left is to replace the block
	# terminator with an explicit `unreachable` marker.
	if block_end > idx

.