This PR refactors functions related to GetXlaTensor(), improving error message, and propagating status. These changes should make errors more actionable from users perspective. Additionally, this PR also allows Python API bound functions to returns status values.

Key Changes:

• Add GetValueOrThrow(Status) overload, so that we can call GetValueOrThrow for both Status ans StatusOr<T> return values
• Properly handle status return types when binding Python functions, i.e. allow calls likescope.def(..., [](...) -> absl::Status { ... })
  • Calls GetValueOrThrow() in the resulting status, returning void
  • Lambda source information correctly appear in the status propagation trace
• Modify error messages of:
  • GetXlaTensorImpl(): assumes less context, with a more detailed error message
  • GetXlaTensor(): more meaningful, and higher level message
  • GetXlaTensors(): more information, such as the index of the non-XLA tensor
  • ReplaceXlaTensor(): reworded
• Create GetInputXlaTensor() for returning errors with more context
  • This should be used instead of calling GetXlaTensor(), specifically for inputs that directly come from the user
• Refactor 2 Python API implementation, as example:
  • _xla_dot_general: calling GetInputXlaTensor() on its tensor inputs
  • _get_graph_hash: calling GetXlaTensors() instead of an explicit loop with the same semantic

Update (Aug 4):

• Created BuildStatusErrorMessage(const absl::Status&), in order to get the error message, and status propagation trace when crashing on a ABSL_CHECK() call.
  • Leaving the C++ stacktrace for future work
• Removed test for _get_graph_hash, since it crashes, now
• Added better error messages for ReplaceTensor functions