Fix corner cases and short-circuit mapreduce

[jishnub]

Addresses certain cases (min/max, and possibly some others) from #301

After this,

julia> C = Fill(1,10000000)
10000000-element Fill{Int64}, with entries equal to 1

julia> @btime mapreduce(identity, max, $(Ref(C))[])
  2.950 ns (0 allocations: 0 bytes)
1

Also fixes certain corner cases to make the reductions consistent with those for Arrays.

cc: @putianyi889

[codecov]

Codecov Report

Merging #302 (831933f) into master (fef31cd) will increase coverage by 0.00%.
Report is 5 commits behind head on master.
The diff coverage is 100.00%.

@@           Coverage Diff           @@
##           master     #302   +/-   ##
=======================================
  Coverage   99.88%   99.89%           
=======================================
  Files           7        7           
  Lines         871      940   +69     
=======================================
+ Hits          870      939   +69     
  Misses          1        1           

Files	Coverage Δ
src/FillArrays.jl	100.00% <100.00%> (ø)
src/fillalgebra.jl	100.00% <100.00%> (ø)
src/fillbroadcast.jl	100.00% <100.00%> (ø)
src/oneelement.jl	100.00% <100.00%> (ø)

📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more

[putianyi889]

Can you use === to check if the types are consistent?

[jishnub]

Looks like InfiniteLinearAlgebra is depending on the incorrect behaviour through BlockBandedMatrices, perhaps inadvertently.

[dlfivefifty]

It looks like downstream breakage is extensive. I believe you that these packages might be depending on buggy behaviour but we should fix them before merging this.