## What changes were proposed in this pull request?
This PR tries to fix the performance regression introduced by SPARK-21517.
In our production job, we performed many parallel computations, with high possibility, some task could be scheduled to a host-2 where it needs to read the cache block data from host-1. Often, this big transfer makes the cluster suffer time out issue (it will retry 3 times, each with 120s timeout, and then do recompute to put the cache block into the local MemoryStore).
The root cause is that we don't do `consolidateIfNeeded` anymore as we are using
```
Unpooled.wrappedBuffer(chunks.length, getChunks(): _*)
```
in ChunkedByteBuffer. If we have many small chunks, it could cause the `buf.notBuffer(...)` have very bad performance in the case that we have to call `copyByteBuf(...)` many times.
## How was this patch tested?
Existing unit tests and also test in production
Author: Wenbo Zhao <wzhao@twosigma.com>
Closes#21593 from WenboZhao/spark-24578.
3f4bda7289