Let empty arrays alias nothing on Julia 1.10

[maleadt]

Julia 1.10's Base.mightalias doesn't check for empty arrays; 1.11 and later do. GPU back-ends like AMDGPU give every empty array the same null data pointer, so on Julia 1.10 reductions, scans and workspace checks rejected empty destinations and sources as aliasing. This caused 221 errors in AMDGPU's GPUArrays testsuite on 1.10, e.g. cumsum(ROCArray(Int[])) and sum(A; dims=2) for a 0×3 A.

Uses an isempty-guarded _mightalias at the three call sites. The tests use reshaped empty views, which reproduce the false positive on CPU arrays on Julia 1.10.

Bumps the version to 0.5.3.

[maleadt]

Superseded: the problem is AMDGPU's pointer-based Base.dataids, which made every empty ROCArray alias every other (and missed overlaps between offset views). AMDGPU's GPUArrays 12 port now drops it in favour of GPUArrays' allocation-based dataids (JuliaGPU/AMDGPU.jl#1145, 1f34fe1a), so AK can keep using Base.mightalias.