Add DreamerV3 implementation

[ParamThakkar123]

Summary

Adds a complete DreamerV3 implementation to torchwm, including:

• DreamerV3 agent (world_models/models/dreamer_v3.py) - Full agent with world model, actor, and critic learned entirely in latent space
• Categorical RSSM (world_models/models/categorical_rssm.py) - Recurrent state-space model with categorical latents
• Block GRU (world_models/layers/block_gru.py) - Block-diagonal GRU layer used in DreamerV3
• Laprop optimizer (world_models/optim/laprop.py) - Loss-adaptive optimizer for stabilizing DreamerV3 training
• DreamerV3 memory (world_models/memory/dreamer_v3_memory.py) - Episodic replay buffer with fixed-size memory
• Vision networks (world_models/vision/dreamer_v3_nets.py) - CNN encoder/decoder tailored for DreamerV3
• DreamerV3 config (world_models/configs/dreamer_v3_config.py) - Structured config with all hyperparameters
• Utility functions (world_models/utils/dreamer_v3_utils.py) - Symlog, OneHotDist, and other helpers
• Tests (~1800 lines across 6 test files) for all components
• Example (examples/dreamer_v3_example.py) - End-to-end training example
• Documentation (docs/source/dreamer_v3.md) - Full API reference and usage guide
• Updated public API - Exports DreamerV3, CategoricalRSSM, DreamerV3Memory, Laprop in the package namespace