This PR deduplicates GetXlaTensors().
Previously, there were 2 functions in the repository with the same name, similar functionality, different signature:

init_python_bindings.cpp

std::vector<XLATensorPtr> GetXlaTensors(const std::vector<at::Tensor>& tensors,
                                        bool want_all);

aten_xla_bridge.h

absl::StatusOr<std::vector<absl_nonnull XLATensorPtr>> GetXlaTensors(
    const at::ITensorListRef& tensors);

• if want_all=true, it behaved similarly to GetXlaTensors()
• otherwise, it collected all XLA tensors from the given list of tensors
• the function in init_python_bindings.cpp lived under an anonymous namespace, not used anywhere else

Therefore, this PR introduces the following key changes:

• Replaces all init_python_bindings.cpp GetXlaTensors(..., /* want_all= */ true) call sites with GetValueOrThrow(bridge::GetXlaTensors(...))
• Rename init_python_bindings.cpp GetXlaTensors() to CollectXlaTensors(), avoiding functions with different semantics to have the same name
• Re-define CollectXlaTensors() to have only the want_all=false behavior, removing want_all flag from the function parameters