* Add new message type `AnnotationProto`
   (`gematria/proto/annotation.proto`).
 * Include it `CanonicalizedInstructionProto`.

 * Add new `Annotation` class.
 * Add `instruction_annotations` vector to `Instruction`.
 * Add Python bindings for these.
 * Add tests for all above changes.

 * Add `Annotation` conversion and update `Instruction` conversion.
 * Refactor `ToVector` and `ToRepeatedPtrField` to be generic.
 * Add Python bindings for conversions.
 * Add tests for all changes.

 * Make instruction annotations available to models via the graph
   builder, where they are stored as a vector-of-vector matrix.
 * Add Python bindings for the same.
 * Add tests for the C++ graph builder implementation, and update the
   tests for the Python bindings to be compatible with the changes.

 * Add a test adding basic blocks with annotations to the graph builder.
 * Add `annotated_basic_blocks_with_throughput.pbtxt` as a source for
   annotated basic blocks for tests.
 * Add utilities to get the annotated blocks to `gematria/testing`.

 * The ends of node embedding vectors used by `TokenGraphBuilderModel`s
   for instructions are overwritten by the corresponding instruction
   annotations.
 * A test for `TokenGraphBuilderModel`s consuming instruction
   annotations has been added as well.

 * Add a `--gematria_annotation_names_file` flag to `run_granite_model`
   to pass in a list of instruction annotation names consumed by the
   model.

 * Move annotation-related tensors from `TokenGraphBuilderModel` to
   `GraphBuilderModelBase`.
 * Adjust the order in which annotation-related tensors are created
   relative to other parts of the graph.
 * Refactor `TokenGraphBuilderModelNodeEmbed`.

 * TF 2.15 is the last version to support Keras 2 without any fixes.
 * Until google#132 lands with a fix that lets us keep using Keras 2,
   this pins the TF version to 2.15 for CI.
 * After this lands, google#132 can be updated to also patch the Keras 2
   usage in this PR and then we can bump the TF version up.