fix: support for tensors/numpy being passed to transcribe() of Canary v2

[coezbek]

What does this PR do ?

This PR fixes that tensors/numpy can't be passed to transcribe() if timestamps==True

Changelog

• This PR matches the batch format of audio tensors against the expected data loader format.
• It also disables chunking if audio tensors are passed, because the tensors don't contain cutting info and need to be cut by the caller.

Usage

• You can potentially add a usage example below

# Add a code snippet demonstrating how to use this 

Pre checks:

• Make sure you read and followed Contributor guidelines
• [NA] Did you write any new necessary tests?
• [NA] Did you add or update any necessary documentation?
• Does the PR affect components that are optional to install? (Ex: Numba, Pynini, Apex etc)
  • Reviewer: Does the PR have correct import guards for all optional libraries?

PR Type:

• New Feature
• Bugfix
• Documentation

If you haven't finished some of the above items you can still open "Draft" PR.

Who can review?

@titu1994, @redoctopus, @jbalam-nv, or @okuchaiev

Additional Information

• Fixes Bug: canary-1b-v2 with timestamps got error when processing transcribe_output #14877

[nithinraok]

Hi @coezbek, thanks for the PR.
Could you also add a test to this file that checks timestamps using NumPy or tensor input in addition to the audio file path?
See the test_aed_forced_aligned_timestamps function.
Add a similar function named test_aed_forced_aligned_timestamps_with_tensors.

[nithinraok]

Follow these steps https://github.com/NVIDIA-NeMo/NeMo/pull/14918/checks?check_run_id=52593146372 to sign off commit and also please send PR to main branch instead of 2.5.0

[coezbek]

@nithinraok I just wanted to point out one way to fix it. I am not sure I can put in the work for a comprehensive fix.