Inline start(::CartesianRange)

Or, alternatively: "Look ma! No CartesianRanges!"

This dramatically simplifies the generated code for iteration over CartesianRanges -- in fact, no references to CartesianRange appear in the LLVM IR with this commit. While it does simplify the code in JuliaLang#9080, it does not solve the performance problem there (I see no difference).  It does, however, speed up `copy(::SubArray)` by 1.3 - 1.6x:

```jl
julia> A = sub(reshape(1:5^3,5,5,5), 1:2:5, :, 1:2:5);

julia> @benchmark copy!(similar(A), A) # current master
================ Benchmark Results ========================
     Time per evaluation: 232.69 ns [227.97 ns, 237.42 ns]
Proportion of time in GC: 0.00% [0.00%, 0.00%]
        Memory allocated: 0.00 bytes
   Number of allocations: 0 allocations
       Number of samples: 4301
   Number of evaluations: 120601
         R² of OLS model: 0.953
 Time spent benchmarking: 5.53 s

julia> @benchmark copy!(similar(A), A) # this PR
================ Benchmark Results ========================
     Time per evaluation: 168.91 ns [165.67 ns, 172.14 ns]
Proportion of time in GC: 0.00% [0.00%, 0.00%]
        Memory allocated: 0.00 bytes
   Number of allocations: 0 allocations
       Number of samples: 4601
   Number of evaluations: 160601
         R² of OLS model: 0.955
 Time spent benchmarking: 5.33 s
```

base/multidimensional.jl

@@ -118,7 +118,9 @@ iteratorsize{I}(::Type{CartesianRange{I}}) = Base.HasShape()
     extest = Expr(:||, cmp...)
     inc = [d < N ? :(iter.start[$d]) : :(iter.stop[$N]+1) for d = 1:N]
     exstop = :(CartesianIndex{$N}($(inc...)))
+    meta = Expr(:meta, :inline)
     quote
+        $meta
         $extest ? $exstop : iter.start
     end
 end