When generating the GUID for a PDB, we use 16 random bytes.  This breaks reproducibility.  In order to fix this, we should use a hash of the PDB, but we probably shouldn't do this by default because PDBs can be very large (several gigabytes), and it will slow down link time considerably.  Once we do this, we can start ignoring the age field (or rather, treating GUID + Age as a unique 20 byte hash).

At the same time, we may also want to investigate the implications of setting the time stamp field in the COFF Debug Directories to be equal to this hash as well.  Currently, we set the debug directory timestamps to be the same as the COFF file header timestamps.  There may be no difference if we use a "more accurate" value, but it's worth investigating anyway.