Fix nested broadcasting of BlockedArray (#485)

Fixes #295. That issue seemed to be caused by code in the generic block
broadcasting logic such as
https://github.com/JuliaArrays/BlockArrays.jl/blob/90ddfa96fb0ade3da4f67ac285fcb340350c19c6/src/blockbroadcast.jl#L121-L129
assuming the broadcasting expression is already flat. This PR adds a
call to `Broadcast.flatten` to explicitly flatten the broadcast
expression in the generic block broadcast code.

It appears that at some point `BlockedStyle` broadcasting expressions
were being flattened in this `Broadcast.instantiate` definition:
https://github.com/JuliaArrays/BlockArrays.jl/blob/90ddfa96fb0ade3da4f67ac285fcb340350c19c6/src/blockbroadcast.jl#L199-L202
but that was changed in #193. Reverting that change broke the tests
introduced in #193 so I guess it is not safe to do that.

I have to admit I'm not sure if this is the best place to call
`Broadcast.flatten`, a lot of this broadcasting code is hard to follow.
The place I put it in this PR fixes #295, doesn't break other tests, and
is hopefully least likely to interfere with custom implementations of
block broadcasting in downstream packages.

Author: mtfishman

src/blockbroadcast.jl

@@ -138,9 +138,10 @@ end
 end
 
 function _generic_blockbroadcast_copyto!(dest::AbstractArray,
-                            bc::Broadcasted{<:AbstractBlockStyle{NDims}, <:Any, <:Any, Args}) where {NDims, Args <: Tuple}
+                            bc1::Broadcasted{<:AbstractBlockStyle{NDims}}) where {NDims}
 
-    NArgs = fieldcount(Args)
+    bc = Broadcast.flatten(bc1)
+    NArgs = length(bc.args)
 
     bs = axes(bc)
     if !blockisequal(axes(dest), bs)

test/test_blockbroadcast.jl

@@ -67,10 +67,26 @@ using StaticArrays
         @test A .+ 1 .+ B == Vector(A) .+ 1 .+ B == Vector(A) .+ 1 .+ Matrix(B)
 
         @testset "preserve structure" begin
-             x = BlockedArray(1:6, Fill(3,2))
-             @test x + x isa BlockedVector{Int,<:AbstractRange}
-             @test 2x + x isa BlockedVector{Int,<:AbstractRange}
-             @test 2 .* (x .+ 1) isa BlockedVector{Int,<:AbstractRange}
+            x = BlockedArray(1:6, Fill(3,2))
+            @test x + x isa BlockedVector{Int,<:AbstractRange}
+            @test 2x + x isa BlockedVector{Int,<:AbstractRange}
+            @test 2 .* (x .+ 1) isa BlockedVector{Int,<:AbstractRange}
+        end
+
+        @testset "nested in-place broadcast" begin
+            x = BlockedVector(randn(4), [2, 2])
+            y = BlockedVector(randn(4), [2, 2])
+            dest = copy(x)
+            dest .+= 2 .* y
+            @test dest ≈ x + 2y
+        end
+
+        @testset "0-dim nested in-place broadcast" begin
+            x = BlockedArray(randn(()))
+            y = BlockedArray(randn(()))
+            dest = copy(x)
+            dest .+= 2 .* y
+            @test dest ≈ x + 2y
         end
     end