Expose configurable silence token ID for CTC decoding

Refactor build system: remove submodules and auto-download dependencies

Fix diversity scoring crash and align logic with paper

Fix beam search pruning logic to work in log-space

Fix memory leak in DecoderState destructor by deleting search roots

Fix FST dictionary memory leak in PathTrie destructor

Ensure directory exists for C++ extension

Add submodules

Add missing utf8 dependency

Update for PyTorch 2 and us pyproject.toml to manage dependencies

Copy .gitignore from https://github.com/github/gitignore/blob/main/Python.gitignore

Copy code from https://gitlab.com/sburud/master/-/tree/master/speechlm/ctc-decode