replace keccakf1600 - with significant performance improvements for ref

[tfaoliveira]

comparison basis (left side): https://github.com/formosa-crypto/libjade/tree/bb8569c01597338148800210de5aa2b512f783c6

new cycles for ref: formosa-crypto/libjade@b53d837
new cycles avx2: formosa-crypto/libjade@7046ee9

ref

keypair, 302134 -> 220474
enc    , 351866 -> 267840
dec    , 424384 -> 339360

avx2

keypair, 85284 -> 83438
enc    , 82856 -> 80916
dec    , 83982 -> 82122

[vbgl]

If you split this PR in two halves, I think I can manage to do the proof for the REF implementation (while postponing the proof for the AVX2 version).

[vbgl]

Well, no longer sure about my previous statement…

[vbgl]

I’m done fixing the proofs. I’ve removed a few redundant instructions.

[tfaoliveira]

Hi Vincent, thank you for pushing on this. Do you think it is ready to be merged?

[vbgl]

Yes it is, if you don’t complain about the changes I’ve made to the code.

[tfaoliveira]

I don't complain. At some point, maybe during the next retreat, I would like to discuss compiler warnings with you and the compiler team.

[vbgl]

Thanks. There was no need to squash the three changes into a single commit.