This closes #2286

This closes #2278.

Savepoints were previously persisted without any meta data using default
Java serialization of `CompletedCheckpoint`. This commit introduces a
savepoint interface with version-specific serializers and stores
savepoints with meta data.

Savepoints expose a version number and a Collection<TaskState> for
savepoint restore.

Currently, there is only one savepoint version:

SavepointV0 (Flink 1.1): This is the current savepoint version, which
holds a reference to the Checkpoint task state collection, but is
serialized with a custom serializater not relying on default Java
serialization. Therefore, it should not happen again that we need to
stick to certain classes in future Flink versions.

The savepoints are stored in `FsSavepointStore` with the following
format:

MagicNumber SavepointVersion Savepoint
  - MagicNumber => int
  - SavepointVersion => int (returned by Savepoint#getVersion())
  - Savepoint => bytes (serialized via version-specific SavepointSerializer)

The header is minimal (magic number, version). All savepoint-specific
meta data can be moved to the savepoint itself. This is also were we
would have to add new meta data in future versions, allowing us to
differentiate between different savepoint versions when we change the
serialization stack.

All savepoint related classes have been moved from checkpoint to a new
sub package `checkpoint.savepoint`.

This closes #2194.

This closes #2162.

The default maximum akka transfer size is 10 MB. This commit reduces the
number of generator blocks from 2^20 to 2^15 which removes the limit on
graph size. The original limit of one millions blocks was intended to
future-proof scalability.

This is a temporary fix as graph generation will be reworked in FLINK-3997.

This closes #2233

This closes #2280.

This closes #2292.

This closes #2283.

The `BlobServer` acts as a local cache for uploaded BLOBs. The life-cycle of
each BLOB is bound to the life-cycle of the `BlobServer`. If the BlobServer
shuts down (on JobManager shut down), all local files will be removed.

With HA, BLOBs are persisted to another file system (e.g. HDFS) via the
`BlobStore` in order to have BLOBs available after a JobManager failure (or
shut down). These BLOBs are only allowed to be removed when the job that
requires them enters a globally terminal state (`FINISHED`, `CANCELLED`,
`FAILED`).

This commit removes the `BlobStore` clean up call from the `BlobServer`
shutdown. The `BlobStore` files will only be cleaned up via the
`BlobLibraryCacheManager`'s' clean up task (periodically or on
BlobLibraryCacheManager shutdown). This means that there is a chance that
BLOBs will linger around after the job has terminated, if the job manager
fails before the clean up.

This closes #2256.

Prevents the CheckpointCommitter from failing a job, if either
commitCheckpoint() or isCheckpointCommitter() failed. Instead, we will
try again on the next notify().

This closes #2287

The user must now specify the ID type as "integer" or "string" when
reading a graph from a CSV file.

This closes #2250

This closes #2248

Suspended jobs were leading to shutdown of the checkpoint coordinator
and hence removal of checkpoint state. For standalone recovery mode
this is OK as no state can be recovered anyways (unchanged in this PR).
For HA though this lead to removal of checkpoint state, which we
actually want to keep for recovery.

We have the following behaviour now:

-----------+------------+-------------------
           | Standalone | High Availability
-----------+------------+-------------------
 SUSPENDED |  Discard   |       Keep
-----------+------------+-------------------
 FINISHED/ |  Discard   |     Discard
 FAILED/   |            |
 CANCELED  |            |
-----------+------------+-------------------

This closes #2276.

This closes #2284.

This closes #2279

This PR adds a JM metric which shows the time it took to restart a job. The time is
measured between entering the JobStatus.RESTARTING and reaching the JobStatus.RUNNING
state. During this time, the restarting time is continuously updated. The metric only
shows the time for the last restart attempt. The metric is published in the job metric
group under the name of "restartingTime".

This closes #2271.

Parameters of some input formats that was only possible to be
set through the Configuration object now have setter methods
that allow the user to do so.

Values set by the setters cannot be reset by the configuration
object.

This closes #2238.

This closes #2120.

Integrated PR comments

This closes #2259.

This closes #2249

- change CLI message upon cluster retrieval
- save JobExecutionResult for interactive executions
- only print Collection size in accumulator results
- remove unused helper method

This closes #2264

This closes #2221.

This closes #2182.

This closes #2078.

This closes #2268

Disposing savepoints via the JobManager fails for state handles or descriptors,
which contain user classes (for example custom folding state or RocksDB handles).

With this change, the user can provide the job JAR when disposing a savepoint in
order to use the user code class loader of that job. The JAR is optional, hence
not breaking the CLI API.

This closes #2083.