[medgan-pr-prep] T7: Update docstrings in medgan.py (Google style)

[jalengg]

Goal

Update all docstrings in pyhealth/models/generators/medgan.py to Google/PyHealth style, matching the pattern established by HALO and CorGAN.

Classes to document

MedGAN (main class)

class MedGAN(BaseModel):
    """GAN-based synthetic EHR generator using a two-phase training strategy.

    Implements MedGAN (Choi et al., 2017) adapted for PyHealth 2.0.
    Pretrains a linear autoencoder before adversarial training with
    standard binary cross-entropy loss.

    Args:
        dataset (SampleDataset): PyHealth sample dataset with ``multi_hot``
            schema for the ``"visits"`` key.
        latent_dim (int): Noise dimension for the generator. Default: 128.
        hidden_dim (int): Hidden dimension for generator residual layers. Default: 128.
        autoencoder_hidden_dim (int): Encoder/decoder bottleneck size. Default: 128.
        discriminator_hidden_dim (int): First hidden layer of discriminator. Default: 256.
        minibatch_averaging (bool): Concatenate batch-mean features in discriminator. Default: True.
        save_dir (str): Directory for checkpoint files. Default: ``"./save/"``.

    Examples:
        >>> from pyhealth.datasets.sample_dataset import InMemorySampleDataset
        >>> samples = [{"patient_id": "p1", "visits": ["401.9", "250.00"]}]
        >>> dataset = InMemorySampleDataset(samples, {"visits": "multi_hot"}, {})
        >>> model = MedGAN(dataset, latent_dim=16, hidden_dim=16, autoencoder_hidden_dim=16)
        >>> isinstance(model, MedGAN)
        True
    """

train_model()

Document both phases, all parameters, return type.

synthesize_dataset()

Returns:
    list of dict: One record per synthetic patient. Each dict has:
        ``"patient_id"`` (str): Identifier, e.g. ``"synthetic_0"``.
        ``"visits"`` (list of str): Decoded ICD code strings.

Sub-modules: MedGANAutoencoder, MedGANGenerator, MedGANDiscriminator

Keep existing structure, add proper Args/Returns sections.

Acceptance criteria

• All public methods have Google-style docstrings
• Class-level docstring includes Args: and Examples: sections
• synthesize_dataset() Returns: block uses named sub-fields (patient_id, visits)
• No undocumented public methods