Implement `ht.broadcast_shapes`

[ClaudiaComito]

Related #891

The heat function stride_tricks.broadcast_shape differs unnecessarily from its numpy and torch counterparts:

1. naming (heat broadcast_shape vs. numpy/torch broadcast_shapes)
2. calling (ht.stride_tricks.broadcast_shape() vs. np.broadcast_shapes() or torch.broadcast_shapes())
3. arguments: ht.stride_tricks.broadcast_shape only takes 2 shapes as arguments, no limit on the numpy/torch side

Feature functionality

ht.broadcast_shapes should behave exactly like np.broadcast_shapes, i.e. return a shape as tuple - while using torch.broadcast_shapes under the hood for GPU compatibility

ADDENDUM: broadcast_shapes is only available starting torch 1.8.0 and numpy 1.20. Doh!