Extremely poor performance of auto global variable distribution with Diagonals

A MWE (1.2 or 1.3RC4):

```julia
using LinearAlgebra
using Distributed
using BenchmarkTools
addprocs(1)
n = 128^2
d = rand(Float32,n)
D = Diagonal(d)

@belapsed @fetch D # ~1 second
```

Looks like its the combination of the wrapping in a `Diagonal` and auto global variable distribution that triggers it, because both of these, which should be functionally equivalent, are fine:

```julia
@belapsed @fetch d # ~10^-3 second
@belapsed let D=D; @fetch D; end # 10^-3 second
```

I'm guessing some of that global variable distribution code is treating Diagonals as dense matrices but should instead be specialized?

Start by running the Julia MWE with Distributed, LinearAlgebra, Diagonal, and @fetch, then compare timings for D, d, and the let-bound D cases. Trace the auto-global distribution path entered by @fetch and inspect how Diagonal is handled versus a dense array. Done means the Diagonal case no longer incurs the reported ~1-second overhead while preserving the existing behavior.

Written by the indexing model from the issue text.