feat: Add model skeleton package with stubs for U-Net, noise schedule…

[KrishanYadav333]

Pre-GSoC groundwork for the "Denoising Astronomical Observations
of Protoplanetary Disks" project (GSoC 2026).

Following up on the config system PR, this lays out the model
module structure before any real implementation starts.
The idea is to define clean interfaces first so the next PRs (actual
forward pass logic, training loop, etc.) have a clear contract
to fill in rather than figuring out the structure mid-way.

All forward methods raise NotImplementedError for now — that's
intentional, not a bug.

What's included:

• UNet — constructor signature with in_channels, base_channels,
  channel_multipliers, attention_levels, dropout
• NoiseScheduler — q_sample, p_sample, p_sample_loop signatures
  with full docstrings explaining what each method should do
• DDPM — wrapper combining UNet + NoiseScheduler, exposes
  training_loss and sample as the two main entry points
• blocks.py — ResidualBlock, AttentionBlock, SinusoidalTimeEmbedding
  stubs with documented args and return shapes
• 20 unit tests covering imports, instantiation, and interface contracts

Files changed:

• src/models/__init__.py
• src/models/blocks.py
• src/models/unet.py
• src/models/noise_scheduler.py
• src/models/ddpm.py
• tests/test_models_skeleton.py

Let me know if the interface design needs any changes before I
start filling in the implementations!