Add distinct left join

[PointKernel]

Description

Contributes to #14948

This PR adds distinct left join. It also cleans up the distinct inner code to use the terms "build" and "probe" consistently instead of "left" and "right".

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[PointKernel]

Benchmark results on RTX8000:

left_join_32bit

[0] Quadro RTX 8000

Key Type	Payload Type	Nullable	Build Table Size	Probe Table Size	Samples	CPU Time	Noise	GPU Time	Noise
I32	I32	0	100000	100000	3984x	129.402 us	13.54%	125.573 us	13.00%
I32	I32	0	100000	400000	2256x	269.767 us	5.79%	266.255 us	5.60%
I32	I32	0	10000000	10000000	30x	16.885 ms	0.23%	16.882 ms	0.22%
I32	I32	0	10000000	40000000	11x	47.211 ms	0.25%	47.207 ms	0.25%
I32	I32	0	10000000	100000000	11x	108.764 ms	0.04%	108.762 ms	0.04%
I32	I32	0	80000000	100000000	11x	159.143 ms	0.04%	159.141 ms	0.04%
I32	I32	0	100000000	100000000	11x	174.163 ms	0.03%	174.160 ms	0.03%
I32	I32	0	10000000	240000000	11x	259.735 ms	0.05%	259.734 ms	0.05%
I32	I32	0	80000000	240000000	11x	299.279 ms	0.02%	299.278 ms	0.02%
I32	I32	0	100000000	240000000	11x	313.821 ms	0.04%	313.821 ms	0.04%

distinct_left_join_32bit

[0] Quadro RTX 8000

Key Type	Payload Type	Nullable	Build Table Size	Probe Table Size	Samples	CPU Time	Noise	GPU Time	Noise
I32	I32	0	100000	100000	10272x	52.113 us	16.87%	48.697 us	14.70%
I32	I32	0	100000	400000	3184x	160.391 us	2.57%	157.164 us	1.39%
I32	I32	0	10000000	10000000	41x	12.482 ms	0.06%	12.478 ms	0.06%
I32	I32	0	10000000	40000000	13x	38.902 ms	0.04%	38.899 ms	0.04%
I32	I32	0	10000000	100000000	11x	89.294 ms	0.03%	89.287 ms	0.03%
I32	I32	0	80000000	100000000	11x	121.486 ms	0.02%	121.482 ms	0.02%
I32	I32	0	100000000	100000000	11x	129.176 ms	0.02%	129.172 ms	0.02%
I32	I32	0	10000000	240000000	11x	212.003 ms	0.05%	211.999 ms	0.05%
I32	I32	0	80000000	240000000	11x	245.199 ms	0.03%	245.195 ms	0.03%
I32	I32	0	100000000	240000000	11x	254.463 ms	0.04%	254.461 ms	0.04%

[bdice]

Do we need tests that explicitly compare distinct join results to "normal" join results? I might have overlooked that, but it would be good to confirm that the results match. (Distinct joins should just be faster since they have an additional requirement on the input.)

[PointKernel]

Do we need tests that explicitly compare distinct join results to "normal" join results?

Distinct join results are always compared against "gold references" instead of the output of normal joins in unit tests. It's a more reliable verification since "normal" joins can go wrong as well.

[ttnghia]

I just have a few questions/nits. Overall LGTM.

[PointKernel]

/merge