Make a variant of `Df.fifo` that is lower latency when it is empty.

`Df.fifo` has 1 cycle latency even when the fifo is empty. We should add new version that can be configured to have 0 or 1 cycle latency when the fifo is empty. The existing `Df.fifo` can then just be hard-configured to have the 1 cycle latency. This will allow the user to make longest-path vs average latency trade-offs.

Start by locating the implementation and uses of `Df.fifo` in the repository. Read how its current empty-FIFO latency is defined and check any related tests or examples before deciding how the variant should be configured. Done means the new FIFO supports selectable 0- or 1-cycle empty latency while existing `Df.fifo` remains fixed at 1 cycle.

Written by the indexing model from the issue text.