b-tree read performance still not good enough

[bnlawrence]

For stupidly large files we're still seeing slow b-tree reads.

On inspection, this occurs, in part, because we do a lot of very tiny reads which we can avoid by reading entire b-tree nodes in one go.

The offending piece of code is this:

 def _read_node(self, offset, node_level):
        """ Return a single node in the b-tree located at a give offset. """
        node = self._read_node_header(offset, node_level)
        keys = []
        addresses = []
        for _ in range(node['entries_used']):
            chunk_size, filter_mask = struct.unpack('<II', self.fh.read(8))
            fmt = '<' + 'Q' * self.dims
            fmt_size = struct.calcsize(fmt)
            chunk_offset = struct.unpack(fmt, self.fh.read(fmt_size))
            chunk_address = struct.unpack('<Q', self.fh.read(8))[0]

There is an obvious optimisation for V1 btrees where we know the length of the entries.