3x regression in broadcast with fused outer constructor

[adienes]

julia> using BenchmarkTools

julia> a, b = [0.1,0.2], [1.0,2.0];

julia> bar2(T, c, f) = sum(f .* T.(c));

julia> @btime bar2(Float32, $a, $b)

# 337.484 ns (6 allocations: 208 bytes) # 1.9
# 1.008 μs (10 allocations: 432 bytes) # 1.10
# 1.100 μs (11 allocations: 432 bytes) #nightly 1.13

found while looking at #39151

I suspect this has also regressed more than once in small increments, but I focused my bisection to the increase in allocations (which was also a 3x in time), which pointed at #49725