Add options to not use bias in Linear (and Conv in future)

[coreylowman]

I think this code be done by implementing Linear as a tuple instead: (MatMul<I, O>, Add<O>), and then if someone doesn't want bias, they can just MatMul<I, O> instead.

However, this would break linear's .npz compatibility with pytorch.

[coreylowman]

Maybe a better option is to include bias as a const BIAS: bool = true generic param.

[jafioti]

I think having a Linear module is much better than just using a tuple. I feel like if you specifically want no bias, then just use MatMul<I, O>

[M1ngXU]

I think having a Linear module is much better than just using a tuple. I feel like if you specifically want no bias, then just use MatMul<I, O>

i think it's something like this
type Linear<I, O> = (MatMul<I, O>, Add<O>)

(and maybe type LinearNoBias<I, O> = MatMul<I, O>)