Implement pairwise distance on 2d grid

[AlexanderKalistratov]

Numbers before fix (size=8192, igpu, float32)

sycl 1d	numba-dpex 2d wrong indexes	numba-mlir 2d wrong indexes
300ms	1s	1s

After fix:

sycl 2d	numba-dpex 2d	numba-mlir 2d
17ms	23ms	17ms

Fixes #290

[adarshyoga]

The faster performance is due to greater parallelism - this kernel is executed with a n*n range whereas the previous version ran on n range for the same input size.
But this precludes us from running on large matrices. For instance, there is a failure with preset=L because the range is greater than the limits set by sycl.

[AlexanderKalistratov]

Blocker IntelPython/numba-dpex#1191

[AlexanderKalistratov]

@adarshyoga @ZzEeKkAa i need an approve to continue

[adarshyoga]

@adarshyoga @ZzEeKkAa i need an approve to continue

@AlexanderKalistratov , I am fine with the changes in this PR for SYCL implementation. But, I don't think we should be changing the numba_dpex_k and numba_mlir_k implementations.
Our goal in dpbench should be to maintain the same logic across kernel and sycl implementations. If there is a performance gap in dpex and mlir implementations because of the way indexing in handled, then we should perhaps be addressing the gap through changes in the frameworks.