map/map! with >2 input arrays is slow

[simonster]

A = randn(100, 100)
B = randn(100, 100)
C = randn(100, 100)
D = randn(100, 100)

type TestFun; end
call(::TestFun, x, y) = x+y
call(::TestFun, x, y, z) = x+y+z

julia> @time map!(TestFun(), A, B, C);
elapsed time: 2.4984e-5 seconds (80 bytes allocated)

julia> @time map!(TestFun(), A, B, C, D);
elapsed time: 0.004941339 seconds (1432016 bytes allocated)

julia> @time map(TestFun(), B, C);
elapsed time: 0.000102774 seconds (80160 bytes allocated)

julia> @time map(TestFun(), B, C, D);
elapsed time: 0.004965889 seconds (1512384 bytes allocated)

This is because map/map! with >2 arguments uses an inefficient strategy to extract the values from the input arrays. Maybe these should be staged functions?

[simonster]

Same issue as #17321