[TF] [stdlib] Improve primitive derivatives and remove some stdlib cruft. (#21943)
* [TF] [stdlib] Improve primitive derivatives and remove some stdlib cruft.
* Improve some derivatives so that they won't capture large tensors.
* Remove some unnecessary `@inline(__always)` in tensor ops. They were required by efficient GPE but are now just burdens on compilation time in optimized builds.
* Remove the `_TFHoistable` and `@autoclosure` hacks in `padded(forSizes:with:)`.
* Update Gradients.swift