Fix float3 alignment bug on Metal for gradient accumulation

[szihs]

The issue was that AtomicTensor used sizeof(T) to compute byte offsets for RWByteAddressBuffer operations, but sizeof(float3) returns 12 while Metal buffers use 16-byte stride for float3.

Changes:

• Add stride-aware atomic methods (atomicAddWithStride, etc.) in atomics.slang
• Add _element_byte_stride field to AtomicTensor in tensor.slang
• Pass element_stride from C++ when binding tensors in slangpytensor.cpp
• Update tests to use proper tensor reads now that alignment is fixed

Fixes #118

[ccummingsNV]

This is a neat and cleanly implemented work around, but it comes with a performance cost both to CPU and GPU, and doesn't help users who want to write atomic accumulators for structs.

@csyonghe @bmillsNV the correct fix here seems to me to be an 'alignedSizeOf' in slang (or simply to make metal return the correct size - depends on what you argue is correct / useful). Is it practical to get that in? I'd rather we do this the right way than with a work around.

If we can agree that is the correct fix, I'm happy to merge tensor work in with that test disabled on Mac, provided we fix it before launching the tensor refactor a few weeks from now.

The implementation of sizeOf does feel a little quirky to me here in general. Coming from other language, I would expect as standard practice to be assume that, for some array, &X[N] == &X[0] + N * sizeof(elementtype(X)). I appreciate that's questioning the definition of size, but in terms of what's useful in practical terms, the stride based definition is the more useful one in my opinion.

[ccummingsNV]

@szihs so I think the correct fix here in the long run is https://discord.com/channels/1303735196696445038/1461282795262316605/1461287168625344671

However I'm happy for this workaround to happen for now, provided it is metal platform only. There's no point paying the cost of extra calculations on other platforms where its not necessary.

If we use the above code, the atomicAdd functions should be tweaked so they simply call atomicAddWithStride, passing in sizeof(T) as the stride.

[ccummingsNV]

Valid work around in principle, but we should make it Metal only and make a few tweaks as documented.

updated. can you take a look

[ccummingsNV]

Happy with the work, though it probably needs a bit of merging with the latest torch integration, as a lot of torch tensor marshall code is now native