implemented std_grad in backward.mojo

[ManishAradwad]

@TilliFe @andresnowak please verify the implementation #9

[andresnowak]

From what i understand, the only thing that is incorrect is this a.grad.simd_load[_nelts](idx_a) * std_derivativeinstead of a multiplication it should be a sum, we have to accumulate the gradient in the a Tensor and doing the chain rule in this operation std_derivative * b.grad.load(idx_b)

[TilliFe]

Thanks @ManishAradwad :). I will merge this as soon as I am back. Leave this PR open.

[ManishAradwad]

From what i understand, the only thing that is incorrect is this a.grad.simd_load[_nelts](idx_a) * std_derivativeinstead of a multiplication it should be a sum, we have to accumulate the gradient in the a Tensor and doing the chain rule in this operation std_derivative * b.grad.load(idx_b)

Right! I'll correct it