V2.0 update

[kaylode]

Detailed Changes:

• Logger Overhaul:

  • Fixed LoggerObserver singleton pattern for thread-safety and session persistence.
  • Added master rank check for distributed training in LoggerObserver.log to prevent duplicate logs across GPUs.
  • Improved LoggerObserver.text() to support multiple args and dictionary-to-JSON serialization.
  • Resolved log duplication by making loguru handlers (StdoutLogger/FileLogger) instance-specific using name filters.

• Subpackage Decoupling:

  • Moved heavy domain-specific dependencies (lightning, torchvision, transformers) to optional extras.
  • Converted theseus/ and theseus/base/ init files to lazy-loadable structures (removed eager wildcards).
  • Implemented lazy imports in LoggerObserver to break global dependencies on PyTorch/Plotly.

• ML Module Simplification:

  • Removed monolithic MLPipeline, MLTrainer, and custom ML-specific callbacks/metrics.
  • Replaced with streamlined tradml.py containing direct-fit functions and TradMLTuner with Optuna integration.
  • Fixed critical bugs in LabelEncode (lazy column initialization) and FillNaN (None-value handling).

• CI/CD & Infrastructure:

  • Updated pyproject.toml and uv.lock to modernized dependency stack.
  • Rewrote tabular test suite in tests/tabular/ to use the new simplified functional API.
  • Preserved existing preprocessors, reduction, and visualization utilities.

• Bump version to 2.0.0

• Modernize all dependencies (Lightning 2.4+, wandb 0.17+, optuna 3.6+, etc.)

• Add huggingface-hub and safetensors as core dependencies

• Replace black/isort with ruff for linting and formatting

• Modernize Dockerfile (CUDA 12.4, Ubuntu 22.04, uv)

• Enhance Registry with generics, merge(), get_or_none(), len, getitem

• Refactor BasePipeline with shared _PipelineBase, extracted helpers

• Fix LightningModelWrapper autocast device detection, use optimizers() API

• Refactor LoggerObserver with dispatch table (O(1) routing)

• Make Metric an ABC with @AbstractMethod decorators

• Cache inspect.signature in getter.py via lru_cache

• Add FSDP strategy support to Trainer

• Add HuggingFaceHubMixin (save/load/push with safetensors)

• Add HuggingFaceHubCallback for automatic model pushing

• Modernize all GitHub workflows (actions v4/v5, Python 3.11, uv, caching)

• Add lint.yml workflow (ruff check + format)

• Add release.yml workflow (PyPI publishing via trusted OIDC)

• Create AGENT.md development guide