Use larger block size for default splice implementation
I saw some glib using code with g_output_stream_splice() being a bit slow today, an I saw that g_output_stream_real_splice() hardcodes a block size of 8kb.
According to https://eklitzke.org/efficient-file-copying-on-linux and the linked to http://git.savannah.gnu.org/cgit/coreutils.git/tree/src/ioblksize.h cp uses 128kb which seems to produce better transfer rates.
Many splices are pretty small, so we should perhaps copy a few 8k blocks first using the on-stack buffer, and then malloc a 128k buffer (in case the thread stack is small).