Progress tracker for all compatibility issues with Python Array API specification:

• __array_namespace__ (add at/towards the end, it's the attribute that declares compliance)
• Support for uint16, uint32, and uint64 #58734
• type promotion with 0d-tensors diverges from array API specification #58736
• Python Array API: support constants #58739
• support for scalars in logical_(and|not|or|xor) #58740
• torch.size() diverges from array API specification #58741
• Python Array API New Operators Tracker #58742
• Implement improved DLPack support #55090
• torch.(min|max)(..., dim=...) diverges from array API specification #58745
• torch.nonzero diverges from the specification (it returns a tensor, not a tuple, by default) - see [numpy] torch.nonzero is similar to np.argwhere not np.nonzero #64502
• Support negative step sizes for slicing #59786
• Remove support for multiple ellipses in slicing #59787
• Bitwise shifts should retain the same dtype as the first input #59867
• Python scalars should be promoted to the same dtype as the respective tensor #59868
• torch.asarray does not detect dtype of Python scalars #70591
• full should take an integer size #70906
• finfo(...).tiny should be aliased to finfo(...).smallest_normal #70909
• eye should support other diagonals than the main one #70910
• support setting step in arange without setting end #70914
• arange should return empty array if bounds are inconsistent with step sign #70915
• uint8 scalar tensors cannot be used for integer indexing #70916
• make torch.(ceil|floor|round|trunc) no-ops for integer inputs #70918
• linspace should support an endpoint parameter #70919
• unique should be split into four partial functions #70920
• sort should only return the sorted input #70921
• argsort is missing the stable parameter #70922
• torch.where to support Python scalars and type promotion #9190
• change supported arguments for parameter dim in squeeze #70924
• allow dim=None in concat #70925

cc @ezyang @gchanan @zou3519 @bdhirsh @jbschlosser @anjali411