effects: replaces usages of ALWAYS_FALSE with TRISTATE_UNKNOWN

[aviatesk]

Although they are basically equivalent in the current implementation,
it would be more correct conceptually if we propagate TRISTATE_UNKNOWN
instead of ALWAYS_TRUE, as it is hard or impossible to derive a global
conclusion from a local analysis on each statement in most places.

This commit also adds a docstring that simply explains the design of the
effect analysis.

[Keno]

I'm ok with this, though we may have to undo it and actually make use of this properly in a few weeks, but until then this is probably cleaner.

[aviatesk]

actually make use of this properly in a few weeks

Well, how do we really want to use it? It sounds a bit tricky to make a global conclusion from local analysis. Maybe use a sort of simple domination analysis like linear_inline_eligible?:

julia/base/compiler/ssair/inlining.jl

Lines 1457 to 1463 in 3888176

	function linear_inline_eligible(ir::IRCode)
	length(ir.cfg.blocks) == 1 || return false
	terminator = ir[SSAValue(last(ir.cfg.blocks[1].stmts))][:inst]
	isa(terminator, ReturnNode) || return false
	isdefined(terminator, :val) || return false
	return true
	end