swift

0902283f - [TF] Remove `unbroadcast(to:)` and improve derivative performance.

[TF] Remove `unbroadcast(to:)` and improve derivative performance. In the pullback for operators that broadcast, use `Raw.broadcastGradientArgs(s0:s1:)` to compute reduction indices instead of using the inefficient `unbroadcast(to:)`. `unbroadcast(to:)` was introduced only for defining derivatives for broadcasting operators and has no practical use, so now we remove it. Operators affected: - `Tensor.+(_:_:)` - `Tensor.-(_:_:)` - `Tensor.*(_:_:)` - `Tensor./(_:_:)` - `min(_:_:)` - `max(_:_:)` - `pow(_:_:)`