Based on a patch by Roderick Colenbrander.

Instead just revert the position slightly, this is a lot easier to maintain.

After a flush call there should be no more samples in the buffer.

This reverts commit 1ba8ece7.

Code for holding sample doesn't make sense and shouldn't be there
because it should really be handled by keeping track of the last
successful position of a frame sent and aborting the whole processing
as soon as rejection occurs.

advapi32: Make sure not to return a credential with a NULL UserName field when a Mac Keychain doesn't have an account name attribute.

dbghelp: Revert commit ae1791b0 and correctly fix what this patch was trying to do.

Otherwise on Windows longjmp may want to mess with the exception frame.