Currently, when seeking, we always discard the BlockReader and reconnect at the
new offset. This is often wasteful, but especially when seeking forward would
keep us in the same chunk, since on reconnecting we discard the beginning of the
chunk anyway.

This adds a new method to BlockReader, Skip, which attempts to discard bytes in
order to skip forward, and makes use of that for small (<64k) forward seeks.

Fixes #275

This reverts commit 5298db4.

Fixes #219

Fixes #244.

…r.Close

Previously, we would simply ignore the result field on CompleteResponse, which
lead to a silently hanging lease on the file. A hanging lease isn't that bad -
it expires on its own - but it was causing the sneaky
AlreadyBeingCreatedExceptions we were seeing in append tests.

Exposing the error (and adding a function IsErrReplicating to unwrap the
PathError, for convenience) allows the client to either ignore it, and let the
lease expire on its own, or to retry in a loop, as the java client does, until
the file is complete and the lease is destroyed.

This has the potential to be a breaking change, since we would silently swallow
the error before.

Fixes #73.