Testing nbdcopy with nbdkit-sparse-random-plugin as a test harness
under perf showed that 34% of the total time was taken calling
read_block() to generate random data within each block (about 18% when
reading and another 16% when writing and verifying).  While we could
probably optimize this a bit, it's all pointless make-work when
testing.  As long as there is some non-zero data in each block it's
still a valid test of nbdcopy.

Therefore add a new flag (random-content=true|false) to enable random
content inside each block.  The new default is random-content=false
which means each block has the same random non-zero byte repeated
across the whole block, which is fast.  To get the old behaviour use
random-content=true.

Total time spent running read_block() went from 34% down to 7%.
Total time spent running nbdkit went from 58% down to 37%.