…s and back

- DAGScheduler:
  - For small enough common task data (RDD + closure) send inline with the Task instead of a broadcast
  - Transiently store task binary data in Stage to re-use if possible
  - Compress the common task bytes to save on network cost
- Task: New TaskData class to encapsulate task compressed bytes from above, the uncompressed length
  and reference index if TaskData is being read from a separate list (see next comments)
- CoarseGrainedClusterMessage: Added new LaunchTasks message to encapsulate multiple
  Task messages to same executor
- CoarseGrainedSchedulerBackend:
  - Create LaunchTasks by grouping messages in ExecutorTaskGroup per executor
  - Actual TaskData is sent as part of TaskDescription and not the Task to easily
    separate out the common portions in a separate list
  - Send the common TaskData as a separate ArrayBuffer of data with the index into this
    list set in the original task's TaskData
- CoarseGrainedExecutorBackend: Handle LaunchTasks by splitting into individual jobs
- CompressionCodec: added bytes compress/decompress methods for more efficient byte array compression
- Executor:
  - Set the common decompressed task data back into the Task object.
  - Avoid additional serialization of TaskResult just to determine the serialization time.
    Instead now calculate the time inline during serialization write/writeExternal methods
- TaskMetrics: more generic handling for DoubleAccumulator case
- Task: Handling of TaskData during serialization to send a flag to indicate whether
  data is inlined or will be received via broadcast
- ResultTask, ShuffleMapTask: delegate handling of TaskData to parent Task class
- SparkEnv: encapsulate codec creation as a zero-arg function to avoid repeated conf lookups
- SparkContext.clean: avoid checking serializability in case non-default closure serializer is being used
- Test updates for above

This avoids runtime erasure for add/value methods that will result in unnecessary boxing/unboxing overheads.