Define an error handling behavior of Serializer.SerializeBatch function

## Feature Request

Currently, the serializer interface doesn't define error handling behavior of [SerializeBatch function](https://github.com/influxdata/telegraf/blob/master/plugins/serializers/registry.go#L43).

If an error is occurred in the middle of processing multiple metrics, whether of not return the rest of succeeded metrics is up to the implementation.

### Proposal:

Define the expected behaviour and add it in the comment/document.

### Current behavior:

Undefined.

It made each serializer uses different strategy.

For example,

* `influx`: https://github.com/influxdata/telegraf/blob/master/plugins/serializers/influx/influx.go#L114-L122
* Ignore certain type of error and return nil if others
* `graphite`: https://github.com/influxdata/telegraf/blob/master/plugins/serializers/graphite/graphite.go#L102-L111
* returns nil for any single error
* `wavefront`: https://github.com/influxdata/telegraf/blob/master/plugins/serializers/wavefront/wavefront.go#L102-L105
* A little unclear but it looks like it ignores all the errors

### Desired behavior:

Having well defined set of behavior including

* Whether of not stop processing all the metrics if error occurs from one of the metrics
* Whether of not log the error if we ignore the error
* Maybe we can change the return error type to `[]error`
* Possibly having a config flag for changing the behavior

### Use case:

This is important to reduce possible confusions of end users and plugin writers.

Start with the SerializeBatch interface in plugins/serializers/registry.go, then compare the referenced influx, graphite, and wavefront implementations. First resolve the expected behavior for partial errors, stopping, and logging; done means the chosen contract is documented and the serializer strategies no longer diverge from it.

Written by the indexing model from the issue text.