[TF] Remove GPE-related dead code from TensorFlow.

[rxwei]

This PR removes dead code from the TensorFlow library that was initially added only for GPE. For context, GPE was removed by #24418.

• Remove _ExpressibleByTensorFlowOp.
• Remove most @inline(__always) from TensorFlow APIs, which were added in early 2018 when GPE deabstraction was not able to recursively inline all functions that use #tfops.
• Remove TensorSendableReceivable and its conformances, which were added in early 2018 when GPE send/receive was being developed.
• Remove functions and protocol requirements added for scalar promotion. Scalar promotion was a feature introduced in early 2018 that was able to turn code like Tensor(x.sum()) into send/receive-free data flow in a TensorFlow graph. It was already disabled a long time ago. The following are removed:
  • Protocol requirements _getScalarOrDie(_:) and _getScalar(_:) in _TensorFlowDataTypeCompatible
  • Functions _TFGetScalarOrDie(_:) and _TFGetScalar(_:)
• Remove functions and protocol requirements added for code hoisting for partitioning (more specifically, hoisting code before the "tensor start point" in a partitioning region). These were added in early 2018, because GPE at the time did not support send/receive, and that host-only random initializers would require send/receive if they appeared inside a partitioning region. The following are removed:
  • Protocol requirement _hoistableClosure(_:) in _TensorFlowDataTypeCompatible
  • Function _TFHoistable(_:)
• Remove compiler-known functions for constant tensor initialization. They were used for enabling the partitioning pass to transform array literals intro Const nodes in TensorFlow graphs to avoid copies. The following are removed:
  • _TFTensorFromScalars(_:shape:)
  • _TFTensorFromScalars1D(_:)
  • _TFTensorFromScalar(_:)
  • _TFStringTensorFromStrings(_:shape:)
  • _TFStringTensorFromStrings1D(_:)
  • _TFStringTensorFromString(_:)

[rxwei]

@swift-ci please test tensorflow

[lattner]

This is all fine with me.