Using __array_interface__ instead of adding __sycl_usm_array_interface__

[PokhodenkoSA]

Using 2 properties __array_interface__ and __sycl_usm_array_interface__ in dparray requires 2 checks in downstream libraries.
Performance penalty.
Related to #184.

/CC @oleksandr-pavlyk @shssf

[oleksandr-pavlyk]

These two properties serve different purposes, and can not be combined. One does not have to check for both. Please clarify what performance penalty you have in mind.

[PokhodenkoSA]

different purposes

Does it mean that we will always know exactly which property we need?

Please clarify what performance penalty you have in mind.

Double checking both properties. One for SYCL queue (__sycl_usm_array_interface__) and another for fallback to usual __array_interface__ if the first property is not available.

[PokhodenkoSA]

I know this cases from dpNP and @shssf.
Also I have asked @DrTodd13 about this new property.
Just want to know the reasons behind introducing this property and comparing to other variants (i.e. extend __array_interface__ dict with new items).

[oleksandr-pavlyk]

NumPy checks for __array_interface__ trying to recognize a Python object as a "duck-array", also known as "array-like" . This property has been largely superseded by Python buffer protocol.

Python objects backed by USM memory export __sycl_usm_array_interface__ to be recognized as carriers of USM memory.

[diptorupd]

Closing since @oleksandr-pavlyk's comments explains the rationale of why the two attributes are needed.