Refactor codebase: package restructure, pyproject.toml migration, and full internal refactor

[alexwolson]

This pull request consolidates two major refactoring efforts into a single, cohesive update. It modernizes the repository structure and dependency management, while also performing a deep internal refactor to improve correctness, performance, validation, and long-term maintainability.

---

1. Repository & Packaging Modernization

Package structure

• Introduced a proper Python package under neurcam/
  • model.py, loss.py, and __init__.py
• Removed legacy flat files (NeurCAM.py, Loss.py)
• Added py.typed marker to support static type checking

Dependency management

• Replaced environment.yml with pyproject.toml
• Switched to the hatchling build backend
• Defined core runtime dependencies (torch, numpy, pandas, scikit-learn, etc.)
• Added optional dependency groups:
  • dev for development tooling
  • wrds for optional integrations
• Documented CUDA-enabled PyTorch installation paths
• Standardized installation workflows:

pip install -e .
# or
uv pip install -e .

Import path update

# Before
from NeurCAM import NeurCAM

# After
from neurcam import NeurCAM

---

2. Compatibility & Deprecation Fixes

• Removed deprecated verbose parameter from ReduceLROnPlateau (PyTorch 2.x)
• Fixed progress bar method calls when verbose=False
• Added missing type annotation for verbose
• Updated scheduler type hints to non-deprecated forms using forward references (e.g., "torch.optim.lr_scheduler.LRScheduler")

---

3. Full Internal Code Refactor

Correctness fixes

• Moved scheduler.step() to run once per epoch (after the batch loop) instead of once per batch
• Simplified torch.cdist() usage in FuzzyCMeansLoss by removing unnecessary unsqueeze() / squeeze()
• Resolved forward-reference issues in type hints

Validation improvements

• Added parameter validation across constructors:
  • NeurCAM (e.g., k, m, epochs, batch_size, etc.)
  • NeurCAMModel (e.g., input_dim, repr_dim, etc.)
  • FuzzyCMeansLoss (e.g., m)
• Added validation for:
  • kl_weight (must be non-negative)
  • single_feature_channels and pairwise_feature_channels (must be non-negative)
  • PyTorch device specifiers (now accepts any valid value such as "cuda:0", "cuda:1", "mps", "cpu", etc.)

Performance & code quality

• Vectorized _o1_valid_cuts and _o2_valid_cuts to remove loop-based logic and reduce duplication
• Optimized tensor initialization in _forward to avoid unnecessary .to() calls
• Added comprehensive input validation in _prepare_data
• Simplified dictionary membership checks in _separated_forward_o1
• Improved readability of _build_projection_network using extend()
• Applied Black formatting consistently across the codebase

---

4. Status

• All tests pass successfully.
• This PR intentionally combines structural, packaging, and internal refactors to avoid intermediate broken states and to present a clean, modernized baseline for future development.