[VL][FOLLOWUP] Reduce Velox hash shuffle partition buffer memory by evicting large partitions after split

[wankunde]

What changes are proposed in this pull request?

Bug fix for VeloxHashShuffleWriter::doSplit.

How does this issue happen :

• After set splitState_ == kInit, do evictPartitionBuffers() to evict large partitions.
• evictPartitionBuffers(pid) calls assembleBuffers, which slices / copies / serializes buffers and may allocate from the Velox memory pool.
• The allocation can trigger memory arbitration, which calls back into reclaimFixedSize.
• Because evictState_ is still kEvictable and evictPartitionBuffersAfterSpill() is true, the arbitrator enters evictPartitionBuffersMinSize, which may select the same pid we are currently evicting.
• The inner eviction resets partitionBuffers_[*][pid], partitionBinaryAddrs_[*][pid], and partitionBufferBase_[pid] while the outer call is still mid-assembly.

How to fix this issue:

• Do do evictPartitionBuffers() before set splitState_ == kInit

How was this patch tested?

Use exists UT

Was this patch authored or co-authored using generative AI tooling?

NO