r348473 didn't take into account the fact that we use nd_buf for
buffering kernel dump and EKCD headers. This is done before the final
call to netdump_dumper(), which flushes residual data, so we were
clobbering that data. Add a netdump_flush_buf() function to flush
residual data and use it before copying headers into the buffer.

I never noticed this when debugging dumps from a netdump-enabled
syzkaller setup, but it manifested when I enabled zstd compression:
the decompressor would report corruption.