Currently the code copies a struct timespec's raw bits as a pair of
uint64_t. On 64-bit systems this has the same representation, but on
32-bit issues there are two issues:
1. tv_sec is a time_t which is 32-bit on i386 specifically
2. tv_nsec is a long not a 64-bit integer
On i386, this means the assertion should fire as the size doesn't match.
On other 32-bit systems there are 4 bytes of padding after tv_nsec,
which in practice are probably 0, as this data is ultimately coming from
the kernel, so it's deterministic (though the padding bytes are not
required to be preserved by the compiler, so are strictly unspecified).
However, on 32-bit big-endian systems, the padding bytes are in the
wrong half to be harmless, resulting in the nanoseconds being multiplied
by 2^32.
Fix this all by marshalling via a real uint64_t pair like is done by the
real ZFS_ENCODE_TIME.
Fixes: 240afd8c1fcc ("makefs: Add ZFS support")