[VL] Dictionary support in shuffle

[FelixYBW]

Description

one common performance issue of Spark the shuffle. Shuffle data size may impact the performance directly.
While one advantage of columnar shuffle is that we can easily use dictionary based, which is expected to decrease the shuffle data size. Meanwhile the Velox pipeline supports dictionary but now all data is flattened after shuffle. With dictionary shuffle support, the next stage can still use the dictionary data which is expected to save memory as well.

[FelixYBW]

@marin-ma @yikf will work on this.

[FelixYBW]

Doc: https://docs.google.com/document/d/1CT2NWaRxi6qfenzC8pcj3bLGNeuLZyR9KS2fL-kkRlY/edit?tab=t.0#heading=h.a85x4cjt94qd

[zhouyuan]

draft: #9426

[zhouyuan]

CI all passed, Q67 shuffle data reduced

[zhouyuan]

Under performance testing.
Ref: Photon vectorized shuffle
https://www.databricks.com/blog/databricks-sql-accelerates-customer-workloads-5x-just-three-years