Add MLIC series models

[Yiozolm]

Summary

This PR adds the MLIC family of learned image compression models and the multi-context checkerboard latent codecs needed to express their entropy models in CompressAI:

• MLIC
• MLICPlus
• MLICPlusPlus
• MLICv2

The work is related to the following papers:

• MLIC: Multi-Reference Entropy Model for Learned Image Compression
• MLIC++: Linear Complexity Multi-Reference Entropy Modeling for Learned Image Compression
• MLICv2: Enhanced Multi-Reference Entropy Modeling for Learned Image Compression
• Improving inference for neural image compression

Only the MLIC++ implementation is adapted from the official JiangWeibeta/MLIC code.
The MLIC, MLICPlus, and MLICv2 implementations are paper-based reproductions written against CompressAI's model and latent-codec structure.

This PR intentionally does not include the SGA-based inference-time latent refinement used by MLICv2+.
That part is planned for a follow-up PR so this change can focus on the base MLIC-series models and reusable multi-context checkerboard codec components.

Changes

• Add multi-context checkerboard latent codecs.
• Add private MLIC and MLICv2 model helper modules.
• Add MLIC, MLICPlus, MLICPlusPlus, and MLICv2 model classes.
• Register MLIC-series constructors in the image zoo.
• Add an MLIC++ checkpoint conversion example for official upstream checkpoints.
• Add focused tests for codecs, helper layers, model construction, zoo registration, and state-dict loading.

Validation

• uv run ruff check compressai tests examples
• uv run ruff format --check compressai tests examples
• Focused MLIC test suite: 86 passed, 1 skipped