Convert docstrings to NumPy-style across callbacks, loaders, utils and core module

[kozlov721]

Motivation

• Standardize and modernize documentation by replacing legacy epydoc-style docstrings with NumPy-style Parameters/Returns/Raises blocks for better readability and tool support.
• Make function and class documentation consistent across callbacks, loaders, utilities, and the Lightning module to improve IDE hints and Sphinx/auto-doc rendering.

Description

• Rewrote docstrings from epydoc-style @type/@param/@return annotations to NumPy-style sections in many modules including files under luxonis_train/callbacks/, luxonis_train/core/utils/, luxonis_train/lightning/luxonis_lightning.py, luxonis_train/loaders/, and luxonis_train/strategies/ without changing runtime logic.
• Updated method/class summaries and parameter descriptions for callbacks such as ArchiveOnTrainEnd, ExportOnTrainEnd, ConvertOnTrainEnd, EMACallback, GradCamCallback, GPUStatsMonitor, UploadCheckpoint, TestOnTrainEnd, and progress bar implementations to use explicit typed parameter blocks.
• Converted utility and loader docstrings (annotate_utils.py, archive_utils.py, export_utils.py, infer_utils.py, train_utils.py, luxonis_loader_torch.py, luxonis_perlin_loader_torch.py, perlin.py) to NumPy-style and clarified return/raises where appropriate.
• Clarified the public API documentation in LuxonisLightningModule (constructor, forward, export_onnx, setup, checkpoint loading, and various helpers) and added Returns/Raises clauses where applicable, with no changes to behavior.

Testing

• No automated tests were executed as part of this change.

---

Codex Task