Fix DPT decoder bugs for Scenic parity

[bingyic]

Summary

Fixes multiple bugs in pytorch/decoders.py to achieve numerical parity (max diff < 1e-4) with the Scenic/Flax reference implementation.

Changes

1. DPTHead: Add output_activation parameter (default False).
   When True, applies F.relu() after project conv, matching Scenic.

2. DepthDecoder: Replace with classification-based depth prediction:

   • nn.Linear(channels, num_depth_bins) head
   • bin_centers buffer via torch.linspace(min_depth, max_depth, num_depth_bins)
   • Forward: relu(logits) + min_depth → normalize → einsum(probs, bin_centers)

3. ReassembleBlocks: Use F.gelu(x, approximate='tanh') to match JAX default.

4. ConvTranspose kernel: Apply 180° spatial flip during Flax→PyTorch conversion.

5. load_decoder_weights(): Unified weight loading from Scenic .zip checkpoints
   with auto-detection and key remapping for all decoder types:
   pixel_segmentation, pixel_depth_classif, pixel_normals → head

Verification

All three decoder types verified for numerical parity against Scenic reference:

• Normals: max diff < 1e-4 ✅
• Depth (classification, 256 bins): max diff < 1e-4 ✅
• Segmentation: max diff < 1e-4 ✅