### What changes were proposed in this pull request?
This PR proposes a workaround to address the Netty OOM issue (SPARK-24989, SPARK-27991):
Basically, `ShuffleBlockFetcherIterator` would catch the `OutOfDirectMemoryError` from Netty and then set a global flag for the shuffle module. Any pending fetch requests would be deferred if there're in-flight requests until the flag is unset. And the flag will be unset when there's a fetch request succeed.
Note that catching the Netty OOM rather than abort the application is feasible because Netty manage its own memory region (offheap by default) separately. So Netty OOM doesn't mean the memory shortage of Spark.
### Why are the changes needed?
The Netty OOM issue is a very corner case. It usually happens in the large-scale cluster, where a reduce task could fetch shuffle blocks from hundreds of nodes concurrently in a short time. Internally, we found a cluster that has created 260+ clients within 6s before throwing Netty OOM.
Although Spark has configurations, e.g., `spark.reducer.maxReqsInFlight` to tune the number of concurrent requests, it's usually not a easy decision for the user to set a reasonable value regarding the workloads, machine resources, etc. But with this fix, Spark would heal the Netty memory issue itself without any specific configurations.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Added unit tests.
Closes#32287 from Ngone51/SPARK-27991.
Authored-by: yi.wu <yi.wu@databricks.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
00b63c8dc2