Add guide for building and executing using Hamilton API, sans I/O

[jmarshrossney]

I think this separation makes sense:

1. Conduit designed to run a pipeline end to end, including i/o, given a config
2. If you want more fine-grained control, that's fine, just work with the underlying Hamilton API

A how-to guide for this would be useful. One concrete use-case could be parameter optimisation where a pipeline needs to be run many times, varying one or more config inputs, and the outputs used to compute an objective function then discarded instead of writing to disk.