Make Metadata writes atomic

[fako1024]

In order to minimize the likelihood of broken blocks or even full GPDir / GPFile instances we should make the atomic (or at least as atomic as possible). That way we can not only minimize the possibility of corruption but also "recover" even if e.g. all or some of the GPFiles in a GPDir have been written when an issue (e.g. system halt or similar) occurs by simply picking up the file offsets from the last successful write, losing only the block during which the error occurred.

DoD

• Make metadata write atomic (use temporary file, then move after successful serialization)
• Ensure that GPFile instances are opened at their last known offset (not the end) and overwriting is allowed
• Add test(s) to reproduce a failure / recovery scenario