Add support for wav2vec2bert Models

[Korla-tech]

This pull request introduces support for wav2vec2-bert models, adding new conversion scripts, quantization tools, and C++ library targets. The changes include new build targets and scripts for handling wav2vec2-bert models, updates to documentation, and improvements to quantization logic for better compatibility and precision. Additionally, there are enhancements to token handling and rendering in the C++ inference code.

Support for wav2vec2-bert models:

• Added a new Python script models/convert-wav2vec2bert-to-ggml.py to convert HuggingFace wav2vec2-bert models.
• Introduced new C++ library target wav2vec2-bert and associated build logic in src/CMakeLists.txt for handling wav2vec2-bert architectures.

Quantization improvements:

• Added quantize-wav2vec2bert.cpp for quantizing wav2vec2-bert models.
• Improved quantization logic in common-ggml.cpp to only quantize 2D tensors with row sizes compatible with the quantizer block size.

Documentation updates:

• Updated README.md with instructions for converting, running, and quantizing wav2vec2-bert models, including new script and binary names.

Token handling and rendering enhancements:

• Added logic in wav2vec2.cpp for identifying and rendering special tokens (such as space, pad, blank, unk), and improved token rendering for output.
• Added a compatibility wrapper for 1D convolution in wav2vec2.cpp, without this the code crashed on some systems.