Handle sparse BOW token ids

[fallintoplace]

What changed

Vocab::GetBowTokenIds() now sizes its result by the largest non-negative token ID rather than by the number of tokens. That keeps sparse explicit vocab IDs from indexing past the end of the returned vector.

Vocab::Load() now tracks the largest non-negative ID while loading. It preserves existing support for negative non-BOW token IDs, but rejects negative IDs on BOW-prefixed tokens because those cannot be represented in a vector indexed by token ID.

A new C++ regression test covers a sparse BOW token ID at 100, a negative non-BOW ID, and rejection of a negative BOW token ID.

History checked

GetBowTokenIds() was added in 9d15517c8 using id_to_token_.size() as the vector size. That assumes dense IDs, while load_token_ids_from_vocab=True supports explicit IDs.

Tests

• python3 -m py_compile lingvo/core/ops/simple_vocab_test.py
• git diff --check -- lingvo/core/ops/simple_vocab.cc lingvo/core/ops/simple_vocab.h lingvo/core/ops/simple_vocab_cc_test.cc lingvo/core/ops/BUILD
• USE_BAZEL_VERSION=5.3.0 npx -y @bazel/bazelisk test --experimental_repo_remote_exec //lingvo/core/ops:simple_vocab_cc_test does not reach analysis in this checkout because @rules_cc//cc:cc_library.bzl is not declared/resolved by the workspace.
• USE_BAZEL_VERSION=5.3.0 npx -y @bazel/bazelisk test --experimental_repo_remote_exec //lingvo/core/ops:simple_vocab_test fails at the same workspace loading step.