xuning / sherpaonnx

This PR adds support for new streaming ASR models from VOSK (Russian) and Kroko (German) to the sherpa-onnx project. It integrates three new models: one German Kroko streaming model and two Russian VOSK streaming models (regular and int8 quantized versions).

- Adds configuration for three new streaming ASR models (indices 24-26)
- Updates APK generation scripts to include the new models
- Modifies GitHub workflow for model upload and deployment

This PR adds support for the newer version (v3) of NVIDIA's parakeet-tdt-0.6b model by exporting it to sherpa-onnx format. The v3 model supports 25 languages, maintaining the same usage pattern as v2 but with improved language coverage.

Co-authored-by: wangmh <minghu.wang@shenyiai.net>

This PR improves Android APK build process by specifying target ABIs during gradle assembly and fixing arithmetic precision issues in decoder calculations. The changes ensure that APKs are built for specific architectures rather than all architectures at once, and correct integer division operations that could lead to precision loss.

- Adds ABI-specific gradle build parameters to target specific architectures during APK assembly
- Updates APK output paths from outputs/apk to intermediates/apk to match gradle's ABI-filtered build structure
- Fixes integer division precision issues in decoder calculations by using floating-point division

This PR fixes a data type compatibility issue in the GigaAM transducer encoder where the output length tensor could be either int32 or int64, but the code only handled int32. The fix adds runtime type checking and supports both data types.

- Adds runtime detection of encoder output length tensor data type (int32 vs int64)
- Implements conditional data access based on the detected type
- Adds error handling for unsupported data types

…har tokens instead of BPE. (#2423)