Fix _broadcast_in_dim_prim_grad when reduce_dims is empty

[rdspring1]

>>> import jax
>>> import numpy as np
>>> x = np.random.normal(size=(5,))
array([ 0.55060031, -0.03787608, -0.81081705,  0.26335944, -1.4178462 ])
>>> jax.lax.broadcast_in_dim(x, [5], [0])
Array([ 0.5506003 , -0.03787608, -0.81081706,  0.26335943, -1.4178462 ],      dtype=float32)

This ^^^ broadcast_in_dim is a no-op. In the gradient function, the reduce_dims tuple is empty, implying we do not want to remove any dimensions. However, when the dim tuple is empty, torch.sum reduces all dimensions. The solution is to skip the reduction.

[mruberry]

@IvanYashchuk and/or @nikitaved -- would you review this, please?