Fix reinterpret performance

[Keno]

This fixes #25014 by making it more obvious what's going on to LLVM.
Instead of a memcpy loop, we use a new intrinsic that puts an actual
llvm.memcpy into the IR, which is enough for LLVM to fold everything
away. In the benchmark from #25014, we still see some regressions from
0.6, but that is because it needs to dereference through the pointers
in the reinterpret and reshape wrappers. In any real code, that
dereferencing should be loop-invariantly moved out of the inner loop.

[vtjnash]

Glad to see this works. Can we just transparently make this he implementation of ccall(:memcpy) (it looks like we don’t need the extra alignment argument right now)

[tknopp]

@Keno Does this make the alignment workaround

A = Mmap.mmap(fdio(fd), Array{UInt8,1}, prod(dims)*sizeof(T), offset)
reshape(reinterpret(T,A),dims)

as fast as

A = Mmap.mmap(fdio(fd), Array{T,dims}, dims, offset)

under 0.6? I needed to change this here
https://github.com/JuliaIO/HDF5.jl/blob/master/src/HDF5.jl#L1651
in order to make code work that failed in 0.7 due to the alignment issue.

[Keno]

Can we just transparently make this he implementation of ccall(:memcpy)

Ok, let's do it that way. We can always add the aligned version if we need it later.

[Keno]

Does this make the alignment workaround [...] fast

Probably

[StefanKarpinski]

This seems eligible for 1.0.1, right?

[Keno]

Yes, already has the backport label.