Experimenting with alternative siglip loss impl for better dist scaling

[rwightman]

@JeniaJitsev @mehdidc

[JeniaJitsev]

Very nice, giving it a try

[rwightman]

So gather impl here looks like it's a better balance but still scaling issues wrt to throughput. The additions here allow flexiblility to add other impl that can further improve things so will merge what's here for now.