feat: add spleeter_pytorch — TF-free 4/5-stem inference for Python 3.14

[salvatoretrimarchi]

PyTorch re-implementation of the Spleeter U-Net separation pipeline, replacing TensorFlow (unsupported on Python 3.14) with pure PyTorch inference and a hand-written TF1 SSTable checkpoint reader.

Modules added:

• unet.py — U-Net architecture ported from Deezer TF model
  (ELU activations for 5-stems config)
• checkpoint_reader.py — Pure-Python LevelDB SSTable parser for reading
  TF1 .index + .data checkpoint files without TF
• convert_weights.py — Converts Spleeter 4-stem or 5-stem TF1 checkpoint
  weights → PyTorch .pt files (one per instrument)
• separator.py — Complete separation pipeline: STFT (librosa) →
  UNet per instrument → Wiener masking → iSTFT
• init.py — Package entry point

Features:

• Supports 4-stem (vocals/drums/bass/other) and 5-stem (+ piano) configs
• Auto-detects available .pt files; loads only those present
• GPU-accelerated when torch.cuda.is_available()
• No TensorFlow, no Docker, no Python version constraints beyond PyTorch

Pull request title

• I read contributing guideline
• I didn't find a similar pull request already open.
• My PR is related to Spleeter only, not a derivative product (such as Webapplication, or GUI provided by others)

Description

A few sentences describing the overall goals of the pull request's commits.

How this patch was tested

You tested it, right?

• I implemented unit test whicn ran successfully using poetry run pytest tests/
• Code has been formatted using poetry run black spleeter
• Imports has been formatted using `poetry run isort spleeter``

Documentation link and external references

Please provide any info that may help us better understand your code.