Optimize set-like operations

[ttnghia]

Set-like operations such as intersect_distinct and difference_distinct call purge_nonempty_nulls when the input is nullable. This PR optimizes these set APIs by checking the existence of non-empty nulls (using has_nonempty_nulls) before calling to purge_nonempty_nulls.

[ttnghia]

Benchmark comparing before vs after this PR:

# intersect_distinct

## [0] Quadro RTX 6000

|  num_rows  |  depth  |  null_frequency  |   Ref Time |   Ref Noise |   Cmp Time |   Cmp Noise |          Diff |   %Diff |  Status  |
|------------|---------|------------------|------------|-------------|------------|-------------|---------------|---------|----------|
|    2^10    |    1    |        0         | 236.523 us |      14.04% | 235.340 us |      11.01% |     -1.184 us |  -0.50% |   PASS   |
|    2^13    |    1    |        0         | 297.840 us |      14.22% | 294.913 us |       8.53% |     -2.927 us |  -0.98% |   PASS   |
|    2^16    |    1    |        0         | 927.063 us |       4.98% | 928.836 us |       4.78% |      1.774 us |   0.19% |   PASS   |
|    2^10    |    4    |        0         |  25.484 ms |       1.07% |  27.372 ms |       4.52% |      1.888 ms |   7.41% |   FAIL   |
|    2^13    |    4    |        0         | 136.627 ms |       0.81% | 146.370 ms |       0.49% |      9.743 ms |   7.13% |   FAIL   |
|    2^16    |    4    |        0         | 422.196 ms |       0.31% | 466.010 ms |       2.20% |     43.814 ms |  10.38% |   FAIL   |
|    2^10    |    1    |       0.2        | 365.748 us |       9.44% | 346.861 us |      60.68% |    -18.887 us |  -5.16% |   PASS   |
|    2^13    |    1    |       0.2        | 405.400 us |       9.24% | 396.616 us |      45.96% |     -8.784 us |  -2.17% |   PASS   |
|    2^16    |    1    |       0.2        | 833.861 us |       6.33% | 875.925 us |      31.60% |     42.064 us |   5.04% |   PASS   |
|    2^10    |    4    |       0.2        |  24.712 ms |       1.05% |  26.997 ms |       4.53% |      2.285 ms |   9.24% |   FAIL   |
|    2^13    |    4    |       0.2        |  13.221 ms |       1.32% |  11.813 ms |       8.87% |  -1407.219 us | -10.64% |   FAIL   |
|    2^16    |    4    |       0.2        |  95.765 ms |       0.50% |  80.123 ms |       4.85% | -15642.116 us | -16.33% |   FAIL   |
|    2^10    |    1    |       0.8        | 353.305 us |      17.76% | 331.605 us |      51.34% |    -21.700 us |  -6.14% |   PASS   |
|    2^13    |    1    |       0.8        | 368.859 us |       8.63% | 378.645 us |      51.42% |      9.786 us |   2.65% |   PASS   |
|    2^16    |    1    |       0.8        | 392.823 us |      10.70% | 390.992 us |      49.61% |     -1.831 us |  -0.47% |   PASS   |
|    2^10    |    4    |       0.8        |   1.621 ms |       4.38% |   1.504 ms |      27.91% |   -117.088 us |  -7.22% |   FAIL   |
|    2^13    |    4    |       0.8        |   1.748 ms |       3.64% |   1.521 ms |      21.72% |   -226.220 us | -12.94% |   FAIL   |
|    2^16    |    4    |       0.8        |   1.959 ms |       4.83% |   1.929 ms |      29.51% |    -30.942 us |  -1.58% |   PASS   **|**

[PointKernel]

Are the performance regression in several test cases expected?

[ttnghia]

Are the performance regression in several test cases expected?

Yes. Previously, only a trivial check null_count == 0 was used. Now has_nonempty_nulls() is called (if the input is nullable).

But sacrificing the performance regression in those cases can improve more performance in some other cases so this can be considered as an "improvement".

Edit: Such (small) performance regression may be due to noise, as you can see some regression in cases even the input doesn't have nulls.

[vyasr]

I have one open request about getting rid of an unnecessary enum and just using a lambda, but since that got a 👍 I'm approving assuming that change will happen.

[PointKernel]

One non-blocking suggestion.

LGTM

[ttnghia]

/merge