Implemented printing for usm_ndarrays

[ndgrigorian]

• Have you provided a meaningful PR description?
• Have you added a test, reproducer or referred to an issue with a reproducer?
• Have you tested your changes locally for CPU and GPU devices?
• Have you made sure that new changes do not introduce compiler warnings?

Closes #954

Implements __repr__ and __str__ for printing dpctl.tensor.usm_ndarray objects and options for printing.
For the sake of user convenience, dpctl stores its own print options separate from Numpy's.

For arrays large enough to be abbreviated (length > 1000 or whatever the user sets as the threshold), a copy of the nd corners is produced and as little data as possible is transferred to CPU while still leveraging np.array2string. This appeared to be faster than running dpctl.tensor.asnumpy on the array.

Some benchmarks using

q = dpctl.SyclQueue()
z = dpt.arange(10**7, sycl_queue=q)

%%time
repr(z);

CPU times: user 11.3 ms, sys: 765 µs, total: 12.1 ms
Wall time: 13 ms

%%time
q = dpt.asnumpy(z)
repr(q);

CPU times: user 6.18 ms, sys: 62.6 ms, total: 68.8 ms
Wall time: 95.8 ms

%%time
str(z);

CPU times: user 0 ns, sys: 9.63 ms, total: 9.63 ms
Wall time: 14.1 ms

%%time
q = dpt.asnumpy(z)
str(q);

CPU times: user 9.63 ms, sys: 50.3 ms, total: 59.9 ms
Wall time: 64.1 ms

Note that these were each run on new kernels, to avoid caching.

[github-actions]

View rendered docs @ https://intelpython.github.io/dpctl/pulls/1013/index.html

[github-actions]

Array API standard conformance tests for dpctl=0.14.1dev0=py310h8c27c75_9 ran successfully.
Passed: 33
Failed: 801
Skipped: 280

[coveralls]

Coverage increased (+0.2%) to 82.276% when pulling 23b3311 on print-feature into bc9c5be on master.

[github-actions]

Array API standard conformance tests for dpctl=0.14.1dev0=py310h8c27c75_10 ran successfully.
Passed: 33
Failed: 801
Skipped: 280

[github-actions]

Array API standard conformance tests for dpctl=0.14.1dev0=py310h8c27c75_10 ran successfully.
Passed: 33
Failed: 801
Skipped: 280

[oleksandr-pavlyk]

@lgtm! Thank you @ndgrigorian. This is an awesome addition.

Majority of the time of the printing comes from memory transfer:

In [12]: x = dpt.ones(10**7 + 2, dtype='i2', device='gpu')

In [13]: %timeit p._nd_corners(x, 4)
222 µs ± 6.9 µs per loop (mean ± std. dev. of 7 runs, 1,000 loops each)

In [14]: %timeit x.__repr__()
932 µs ± 50.6 µs per loop (mean ± std. dev. of 7 runs, 1,000 loops each)

In [15]: %timeit p._nd_corners(x, 4)
232 µs ± 10.1 µs per loop (mean ± std. dev. of 7 runs, 1,000 loops each)

In [16]: xx = p._nd_corners(x, 4)

In [17]: %timeit dpt.asnumpy(xx)
638 µs ± 10.2 µs per loop (mean ± std. dev. of 7 runs, 1,000 loops each)

Perhaps we can improve the code in the future by avoiding the concat call, and instead allocating the elided memory, and launching several concurrent copy kernels for each kernel, and then a copy to host dependent on those corner copying kernels.

This better be done in a separate PR though.

[github-actions]

Deleted rendered PR docs from intelpython.github.com/dpctl, latest should be updated shortly. 🤞

[github-actions]

Array API standard conformance tests for dpctl=0.14.1dev0=py310h8c27c75_10 ran successfully.
Passed: 33
Failed: 801
Skipped: 280

[fcharras]

TY for this. Was much needed 🙏