[GLUTEN-10660][VL] Adding configuration for hash table build

[zhouyuan]

What changes are proposed in this pull request?

Adding configurations for hash map build optimization in facebookincubator/velox#7066

How was this patch tested?

[liujiayi771]

We need to set spark.gluten.velox.abandonbuild.noduphashminpct=100 to verify the behavior when deduplication is enabled.

[github-actions]

#10660

[liujiayi771]

After the testing, should we enable this feature by default?
In addition, the config key may still change in the upstream Velox. It's better to wait until the code review comments are finalized.

[zhouyuan]

@liujiayi771 I vote for enabling this by default. The staging patch in our Velox fork is not updated. We will update on that after facebookincubator/velox#7066 merged

[FelixYBW]

At least for TPCDS test, I didn't see any performance regression with spark.gluten.velox.abandonbuild.noduphashminpct=100. I think we can enable it by default.

[liujiayi771]

TPCDS has very few queries that contain semi/anti joins, and all of these involve duplicate join keys.

[zjuwangg]

ditto.

[zhouyuan]

@liujiayi771 are you suggesting it's better to disable this feature by default?

[marin-ma]

Perhaps use "percentile" to avoid abbreviations in the doc.

[marin-ma]

LGTM. Thanks!

[liujiayi771]

@liujiayi771 are you suggesting it's better to disable this feature by default?

Yes, for the Q95 build side join keys, the duplication rate is very high, so the optimization is very significant. However, for production jobs without a high duplication rate, there could be a performance regression. It's necessary to adjust the abandon percent based on the specific job conditions.