Reduce memory usage in writer with more memory efficient output buffer implementation

[chenyangfb]

Description

Currently ChunkedSliceOutput is used for storing compressed output in writer. It managed list of buffers with size of power of 2 (e.g. 8k, 16k, 32k), and reuse buffers after flushing. It could leads to extra memory usage and OOM due to 1) mismatch in compressed output size and buffer size, 2) reusing buffers and not freeing buffers leads to extra memory usage by design.

Common scenario which leads to OOM includes

1. large number of streams with small amount of data (100k stream with 1k compressed bytes), each using minimal buffer size (e.g. 8k)
2. Each stream is wasting half of largest buffer (e.g. 8M out of 16M buffer)
3. Writer memory usage is high even after flushing (Reduce memory usage in writer by freeing unused buffers #23724 support freeing unused buffer in chunk supplier during reset)

This PR introduce OrcLazyChunkedOutputBuffer which focus on avoiding used memory.

1. Create buffer size based on the size of compressed output, this avoid the issue 1) and 2) mentioned above
2. lazy initialization in OrcLazyChunkedOutputBuffer and OrcOutputBuffer
3. Reset all the closed buffers, only keep the active buffer.

This behavior is controlled by lazyOutputBuffer in OrcWriterOptions, and it's disabled by default.

Impact

Reduce memory usage in writer.

Test Plan

Tested with Spark workload with high memory usage.
~10% improvement in run time and resource usage (memory reservation time), reduced GC time.
Tested with general Spark workload
No change in cpu time, slight reduction in run time and GC time.

Release Notes

General change
Reduce memory usage in writer with more memory efficient output buffer implementation