The HDFS filesystem should support re-trying blocks with datanode failure

[tarnfeld]

If between the point the client requests a block location from the namenode, the datanode it was given drops out of the cluster, an exception is bubbled to the top... this exception is a result of us not being able to connect to the datanode to get the block.

In this specific case, we should detect a datanode failure and ask the namenode again for a location of the block. I'm not sure how many times we should retry this process, I think the hadoop-client does it indefinitely.

Let's consult the java implementation and follow that.

Note: this filesystem implementation is still sitting in review via #2