Define alias detection for every GPU array through a location hook

[maleadt]

Back-ends each defined their own Base.dataids and Base.mightalias, and they diverge. The aliasing testsuite added in 12.0.3 fails on CUDA, AMDGPU, oneAPI and OpenCL, because an empty contiguous view is reported as aliasing the array around it. Base's check for two SubArrays also converts GPU parents to pointers. That takes stream ownership on CUDA and throws for OpenCL buffers without a device address.

This PR defines alias detection once, for every AbstractGPUArray, on top of a new hook:

GPUArrays.memory_location(A) -> (base::UInt, offset::Int)   # default: nothing

base identifies the allocation (normally its address; a handle works for memory without one) and offset is the byte offset of the first element.

• Base.mightalias: empty arrays alias nothing on every Julia version, including 1.10, which has no isempty check. Two arrays with a location alias when their byte ranges overlap. Isbits-union element types, whose selector bytes lie outside that range, and types without the hook are compared by dataids.
• Base.dataids: the shared DataRef, the allocation's base and the array's start address, or () for empty arrays. Base compares these whenever an array is wrapped (e.g. in a SubArray), so wrapped arrays are compared by allocation, which is conservative.
• Base._parentsmatch: two parents are the same memory if they have the same location, element size and size. GPU arrays are no longer converted to pointers. When parents don't match, Base falls back to dataids, which is always sound.

JLArrays implements the hook in place of its byte-range mightalias from #803. The behaviour for JLArrays is the same, except that isbits-union arrays are now compared by allocation. The docstring documents the contract, including what the hook cannot express: memory reachable at two addresses, and memory wrapped from a pointer into the middle of an allocation when it is compared through a wrapper.

The design was reviewed by two other agents (Codex and Claude). It was checked on Julia 1.10, 1.11, 1.12 and 1.13: the aliasing testsuite for JLArray and Array, JLArray broadcasting, CPU probes of the wrapper, union and mixed GPU/CPU paths, and no new method ambiguities.

This is a minor release: GPUArrays 12.1.0, and JLArrays 0.4.2, which requires it. Once this is merged, the back-ends replace their dataids/mightalias methods with the one-line hook (see NEWS.md).