TransducerRecognizer support

[rdcm]

Hi! I'm currently migrating our speech-to-text service from Python to Rust. Our service relies on the from_transducer API from the original library, and it would be great to have the same API available in sherpa-rs. I still need to add a test, which might take a bit more time. In the meantime, if you have any comments or suggestions regarding the PR, I'm happy to discuss and make changes as needed.

TODO:
- tests (done)

Run tests:

• cargo test transducer

Prerequisites:

• zipformer2 model

[rdcm]

@thewh1teagle test added, pr ready for review

[thewh1teagle]

Hi @rdcm
Can you omit the test and add example instead?
Let me know then if the example works :)

[thewh1teagle]

Follow the other examples template you can copy paste and change a bit

[rdcm]

@thewh1teagle done, example added :)

[thewh1teagle]

Two things I noticed:
(1) when filling sherpa structs, instead of filling every inner struct with nulls we can use mem::zero<_>() it's very cool pattern I used in other files, this way we don't need many structs, the downside is that we need to create the main struct in unsafe block, but it's not big deal imo

(2) in the example instead of filling nulls, you can use ..Default::default() like in other examples

[thewh1teagle]

Feel free to leave it as is, I'll merge it anyway soon : )

[rdcm]

@thewh1teagle

1. Interesting approach, but not used across whole project, for example in whisper.rs and zipformer.rs nulls passed to structs explicitly - done.

2. Done.

Example still works fine :)

[rdcm]

Also added more detailed example.

[thewh1teagle]

Looks great. Thank you!

[rdcm]

@thewh1teagle Hi, do you know if there is any way to pass &[u8] instead of a file path to the model?
Currently, we use a workaround with a memfd_create call to get a file descriptor for the memory and then pass the path to that descriptor.

However, ONNX Runtime provides a more convenient API, for example:

Session::builder()
   ....
   .commit_from_memory(model);