Adapt GPU feature to Supernova / non-folding of running instance

[huitseeker]

This is a consequence of the following upstream issue & PR:

• Optimization: The running instance does not need to absorbed. microsoft/Nova#56
• Remove absorbing of running instance microsoft/Nova#291

This PR will be ported to Arecibo in #276

As a result of not absorbing the running instance, this PR modifies the NUM_FE_FOR_RO constant from 24 to 9, which in turn modifies the IOPattern of the Poseidon circuit:
https://github.com/lurk-lab/arecibo/blob/2858db1da990cadb704d0c20525e40ddd811416e/src/provider/poseidon.rs#L136-L143
This IOPattern parameters are also modified (similarly, 24 ->9) out-of-circuit.

There's a couple of issues with this:

• To this day, we just import a neptune instantiation with arity24 for the GPU, something that may no longer be sufficient,
• SuperNova already modified the # of absorbs we configured the ROs with (see function num_ro_inputs .. the # of absorbs is instance dependent, but we run e.g. simple tests with a # absorbs ranging between 20 and 42), reinforcing the obsolescence of this number (24).

=> should we modify neptune features activated in the cuda feature?

[huitseeker]

Notes from discussion with @porcuquine : Nova and Arecibo aren't actually using GPU acceleration, which in the case of Neptune, can only be gotten through calling into the batch_hasher functions. We can thus remove the Cargo metadata which imports this functionality.

[huitseeker]

Closed in #279