Coverage increased (+0.3%) to 72.733% when pulling 1f34ee8 on scalar_copy into b97efdf on master.

@densmirn can you also implement __index__?

See data-apis/array-api#231

import dpctl.tensor as dpt
X = dpt.usm_ndarray(10, 'd')
float(X[:1])   # raises TypeError

This works with NumPy:

import numpy as np
X = np.ndarray(10, 'd')
float(X[:1])   # produces a float 

It breaks with the current code because Xusm.usm_data gives the container underlying the view which may be bigger than the view:

X = dpt.usm_ndarray(10, 'd')
X[:1].usm_data.copy_to_host().view('d').shape  # outputs (10,)

Thus instead of self.usm_data.copy_to_host().view('d').__float__() we should do self.usm_data.copy_to_host().view(self.dtype)[:0].__float__().

Ideally, in a future, we should avoid copying the entire underlying buffer.

Actually slice of array can be different, e.g.:

import numpy as np
X = np.arange(10)
float(X[:1])  # 0.0
float(X[2:3])  # 2.0

So when we copy data to host we don't know the slice:
self.usm_data.copy_to_host().view(self.dtype)[???].__float__()

Looks like it should be somehow handled on __getitem__ side.

@oleksandr-pavlyk could you please look at last comment?

This is affected by #583

This PR(tests) have uncovered another problem which I fixed.

It was inability of memory objects to understand objects with __sycl_usm_array_interface__ with 'typestr': '|b1'.

Also added tests to cover new functionality.

I also added __complex__ magic (a.k.a. dunder) method.