Bugfix: Use Base.aligned_sizeof instead of sizeof in Mmap.mmap

[JonasIsensee]

The purpose of this PR is to fix #58982 .

The current (bugged) behaviour is:

julia> primitive type MyPrim9 9*8 end

julia> sizeof(MyPrim9)
9

julia> Base.aligned_sizeof(MyPrim9)
16


julia> arr = Vector{MyPrim9}(undef, 2)
2-element Vector{MyPrim9}:
 MyPrim9(0x700000000000000002)
 MyPrim9(0xd000007fb7705a2c30)

julia> write("file", arr)
32

julia> using Mmap

julia> mmap("file", Vector{MyPrim9})
3-element Vector{MyPrim9}:
 MyPrim9(0x700000000000000002)
 MyPrim9(0xd000007fb7705a2c30)
 MyPrim9(0x000000000000000000)

This is because Mmap erroneously uses sizeof instead of Base.aligned_sizeof to compute the number of elements (32 ÷ 9 == 3).
This has gone undetected for a long time because sizeof(T) == Base.aligned_sizeof(T) is true for most types including structs. (Padding is to a multiple of 8 bytes is included in sizeof(T))

closes #58982

edit:
Previously, sizeof would throw an error exception when presented with a non-mappable type such as Ref.
This occurred within the default argument generation. Base.aligned_sizeof does not error here, and this is instead caught in the dedicated isbits check.
This seems desirable and I changed the test to require an ArgumentError.

[JeffBezanson]

Very nice, thank you for finding and fixing this.

[Seelengrab]

Should this be put into a @testset to make failures here more localized?

[JonasIsensee]

We can wrap it into a testset.
I just tried to match the (oldish) style of the test file as closely as possible.

[KristofferC]

I think this broke Windows CI (#59053)

[JonasIsensee]

I think this broke Windows CI (#59053)

The only difference from the other tests is that it is wrapped into an @testset.
Are these known to cause trouble with file removal on windows?

[KristofferC]

CI at least passed on the Windows worker that ran (https://buildkite.com/julialang/julia-master/builds/49176#01980ab7-c042-4ddb-a06e-6020753851d3) before the testset was introduced.