Allow mutation of non-isbits eltypes

[MasonProtter]

For large MArrays, this'll be slower than if we had a proper MutableTuple, but it does work and should be pretty quick for reasonable sizes. It shouldn't screw up the garbage collector or do anything unsafe like the previous version did.

[MasonProtter]

That's so bizarre, this test works totally fine for me locally, and the operations it's doing should be legal.

[serenity4]

This works for me too, perhaps we can give it another try? What were the issues? (logs are no longer available)

I made an attempt as well, with some special-casing for certain mutable types (for which I believe the concerns of #27 do not apply): https://github.com/JuliaArrays/StaticArrays.jl/compare/master...serenity4:setindex-nonisbits. That seems to greatly improve performance for large arrays:

julia> mutable struct Mut; x::Float64; end

julia> @btime setindex!(v, $("ho"), 2) setup = v = @MArray ["ha" for i in 1:100];
  408.465 ns (3 allocations: 1.62 KiB)

julia> @btime setindex!(v, $(Mut(3.0)), 2) setup = v = @MArray [Mut(rand()) for i in 1:100];
  7.032 ns (0 allocations: 0 bytes)

Though this optimization could be implemented later and if we are really 100% sure it won't cause any issues. I'd just try to go forward with the current PR first.