Fix different positional embeddings clashing

[sdtblck]

As mentioned on discord, if using rpe, the regular pos embs were still being added in.

This should fix that - you now specify the type of positional embedding you want as the "--pos-emb" arg and can choose from 'learned', 'sinusoidal' and 'rpe'.

Still needs testing.

[sdtblck]

We might want to double check our sinusoidal implementation, but all the others appear to work fine.

dashed peach line: sinusoidal,
other peach line : learned embedding summed w input,
green : no pos emb
grey : rpe

[StellaAthena]

Are we going to leave this open until the bug in sinusoidal is fixed or would you rather we merged and opened a bug issue @sdtblck

[sdtblck]

Are we going to leave this open until the bug in sinusoidal is fixed or would you rather we merged and opened a bug issue @sdtblck

It's unclear whether it's a bug but if it is, it's in the main branch too. And the rpe stuff just flat out doesn't work in the main branch, so i'd suggest we merge this.

[StellaAthena]

This mostly worlds, however it appears that there’s a bug in the sinusoidal embedding.