wip: move the buffer spec to port

[mormj]

Started down this path, hoping that we could have some default custom buffer for e.g. cuda blocks, but there is no support for the transfer_type which is not known at block instantiation. This needs more thought about what we want the custom buffer api to be.

One benefit of having the custom buffer associated with the port is that we can do something like

graph.connect(x,y)
x.output_port(0).set_custom_buffer(...)

Is that better? I'm not sure.

Signed-off-by: Josh Morman [email protected]