feat: Add U-Net building blocks with shape-consistency tests

[KrishanYadav333]

Summary

Part of pre-GSoC groundwork for the EXXA DDPM denoising pipeline.

Adds the three core reusable blocks that the U-Net encoder/decoder will be built from.

Changes

• SinusoidalTimeEmbedding — log-spaced sin/cos frequency embeddings for diffusion timestep conditioning
• ResidualBlock — GroupNorm + SiLU + Conv2d with time projection injection and 1×1 skip when channels differ
• AttentionBlock — GroupNorm + MultiheadAttention over flattened spatial dims with residual connection

Tests

17 shape-consistency tests in tests/test_blocks.py covering:

• Output shapes and spatial preservation
• Variable batch sizes
• Both skip connection paths (identity and 1×1 projection)
• dtype correctness

All 17 pass.

Next

Next PR will assemble these blocks into the full U-Net encoder/decoder stack.