[copybara-service]

Stop statically linking TensorFlow into every OSS TF.Text op library.

Since TF 2.20 the TensorFlow build enables pywrap rules, so if_static() resolves to full implementations, and TF proto targets are no longer header-only. Because every OSS py_tf_text_library depended on @org_tensorflow//tensorflow/lite/kernels/shim:tf_op_shim, each _<op>.so statically linked its own copy of TF protos, the protobuf runtime and absl (5–9 MB per op library). That inflated the wheel about 8x and put duplicate TensorFlow, protobuf and absl state into the process.

Compile only the three shim sources against the TensorFlow wheel headers and link them dynamically against libtensorflow_framework, like any other custom op. A small patch exports the shim sources from the TensorFlow source tree and keeps the AttrSlice::Find lookup in TfInitContext::GetAttr, which is required because Map::at() on a NodeDef owned by libtensorflow_framework fails across the shared-library boundary.

With that, tensorflow.xla_if_static.bzl.patch no longer affects the op libraries; it only added TF protos to tflite_op_shim, which libpybinds_library_common.so links as a whole archive. Remove the patch together with the unused TensorFlow core includes that still required it: whitespace_tokenizer_kernel_template.h no longer includes tensor.h, sentence_fragmenter_v2.{h,cc} no longer include status.h (the struct uses uint32_t and the test std::string instead of the TF aliases), and fast_wordpiece_tokenizer_model_builder.cc includes absl/log/log.h directly instead of getting LOG transitively.