[runtime] upgrade openfst to v1.8.5 and WeTextProcessing to v1.2.0

[pengzhendong]

Summary

• Upgrade openfst from kkm000/openfst 1.7.2.1 to csukuangfj/openfst v1.8.5
• Pin WeTextProcessing to v1.2.0 with SOURCE_SUBDIR runtime
• Bump gflags to v2.3.0, glog to v0.7.1 (CMake compatibility)
• Remove obsolete openfst patch files
• Add compatibility layer for kaldi/decoder code

Fixes #2816

Details

Why upgrade

WeTextProcessing v1.2.0 changed StringCompiler/StringPrinter construction to use the default TokenType::BYTE parameter. This requires openfst >= 1.8.5 (csukuangfj fork). wenet previously used kkm000/openfst 1.7.2.1 which has no default for the first constructor argument, causing a compile error.

openfst upgrade (openfst.cmake)

Switch from kkm000/openfst 1.7.2.1 to csukuangfj/openfst v1.8.5-2026-04-11. The new fork has a proper root CMakeLists.txt, so FetchContent_MakeAvailable works directly. Removed the old PATCH_COMMAND since the patched CMakeLists.txt/log.h/flags.h files are no longer needed — the new openfst ships its own CMake build system and uses FST_FLAGS_ prefix for flag variables (avoiding gflags conflict).

WeTextProcessing v1.2.0 (wetextprocessing.cmake)

Added SOURCE_SUBDIR runtime so FetchContent_MakeAvailable finds CMakeLists.txt in the runtime/ subdirectory (the repo root has only Python code). Removed manual add_subdirectory calls — the runtime CMakeLists.txt handles them via add_subdirectory(utils) and add_subdirectory(processor).

gflags/glog version bumps (gflags.cmake, glog.cmake)

gflags v2.2.2 uses cmake_minimum_required(VERSION 2.8.12) which is rejected by CMake >= 3.31. Bumped to v2.3.0 and v0.7.1, matching WeTextProcessing v1.2.0's dependency versions.

Removed old openfst patches (6 files, -619 lines)

The patch files (log.h, flags.h, flags.cc, CMakeLists.txt) were designed for kkm000/openfst. They commented out openfst's LOG/CHECK macros and included glog/gflags to avoid naming conflicts. No longer needed with the new openfst.

Compatibility layer

• fst_types.h (new): Type aliases int32, int64, uint64 etc. The old openfst provided these in fst/types.h which no longer exists.
• kaldi-types.h: Include fst_types.h instead of <fst/types.h>.
• kaldi-error.h: Include glog for google::LogMessage (was previously obtained transitively through the patched openfst). Restore DECLARE_* macros after glog's logging.h #undefs them at end of file. These macros must use FST_FLAGS_ prefix for openfst's own internal flags.
• utils/flags.h: Include gflags/glog. Restore DECLARE_* macros same as above. Note: DEFINE_* comes from gflags with FLAGS_ prefix, completely separate — wenet's own flag passing is unaffected.
• utils/CMakeLists.txt: Link gflags_nothreads_static and glog — previously obtained transitively through fst.

kaldi API updates

New openfst renamed several types and uses FST_FLAGS_ prefix for flag variables:

• MapFst→ArcMapFst, MapFstOptions→ArcMapFstOptions, Map()→ArcMap()
• FLAGS_fst_field_separator→FST_FLAGS_fst_field_separator, FLAGS_fst_weight_separator→FST_FLAGS_fst_weight_separator

Build system

• kaldi/CMakeLists.txt: Explicit STATIC for kaldi-util and kaldi-decoder. The new openfst sets BUILD_SHARED_LIBS=ON on non-Windows, causing kaldi-decoder to build as .dylib. But kaldi-decoder depends on ContextGraph::GetNextState from the decoder library (which links back to kaldi-decoder), creating a circular dependency on macOS.

Small fixes

• post_processor.cc: string::npos→std::string::npos (old openfst's using std::string is gone).
• context_graph.cc: Remove streaming CHECK(...) << "msg" syntax — new openfst's CHECK returns void, not a stream.