The new P2PDataStore class was sending full blocks in headers messages,
which meant that calls to send_blocks_and_test() would blow up memory if
called with a large number of blocks. Fix that by only sending headers
in headers messages.

This means that python should use just over 1GB for feature_block.py (with bitcoind also using just over 1GB). That's the same as before the feature_block.py refactor.