Don't create nested thunks when accumulating

Otherwise, it's too easy to create massive types that freeze compilation and blow the stack.

src/lib/lib.jl

@@ -43,9 +43,9 @@ accum(x::ChainRulesCore.Tangent, y::NamedTuple) = accum(wrap_chainrules_output(x
 accum(x::Nothing, y::AbstractThunk) = y
 accum(x::AbstractThunk, y::Nothing) = x
 
-accum(x, y::AbstractThunk) = @thunk(accum(x, unthunk(y)))
-accum(x::AbstractThunk, y) = @thunk(accum(unthunk(x), y))
-accum(x::AbstractThunk, y::AbstractThunk) = @thunk(accum(unthunk(x), unthunk(y)))
+accum(x, y::AbstractThunk) = accum(x, unthunk(y))
+accum(x::AbstractThunk, y) = accum(unthunk(x), y))
+accum(x::AbstractThunk, y::AbstractThunk) = accum(unthunk(x), unthunk(y))
 
 # Core functions
 @_adjoint_keepthunks deepcopy(x) = deepcopy(x), ȳ -> (ȳ,)