[BUG] decompressed batches corrupt if they are made spillable

[abellina]

While working on #7777 I ran into an issue where a decompressed batch (via nvcomp/UCX) was made spillable, but then I got a corrupted batch out when calling getColumnarBatch. This is likely an issue of 23.04.

So far, when we decompress batches in GpuCoalesceBatches https://github.com/NVIDIA/spark-rapids/blob/branch-23.02/sql-plugin/src/main/scala/com/nvidia/spark/rapids/GpuCoalesceBatches.scala#L645, we are taking the TableMeta directly from the compressed vector, instead of building a new one without compression info.

Code that uses the metadata to rebuild the ColumnarBatch would produce an invalid batch, because we do different things when the batch has codecs defined: https://github.com/NVIDIA/spark-rapids/blob/branch-23.02/sql-plugin/src/main/scala/com/nvidia/spark/rapids/RapidsDeviceMemoryStore.scala#L282, and

spark-rapids/sql-plugin/src/main/scala/com/nvidia/spark/rapids/RapidsBufferStore.scala

Line 309 in 96ed06b

GpuCompressedColumnVector.from(devBuffer, meta)

I think I can fix this with #7777, but adding an issue since it's a bug, and I am not sure if it was made worse in 23.04 because of #7572, since we now rely on TableMeta on the first creation of a batch from a RapidsBuffer.

[abellina]

Related issue: #7850