add indice key to basicblock sparse encoder

[GregorioFornetti]

Motivation

I found that when setting block_type="basicblock", the encoder was not passing the indice_key into the SparseBasicBlock generation loop.

This creates a significant computational overhead when using the spconv library. Because the keys default to None, the backend is forced to completely regenerate the spatial index mapping for every single SubMConv3d layer during each forward pass, instead of reusing the maps already created by previous layers in the same stage.

With this modification, I obtained a small performance gain in CenterPoint-Voxel for nuScenes. The FPS in a P100 GPU went from 12.27 to 13.05.

Modification

To fix this problem, I made the following changes:

1. Propagated the Key: Added the indice_key=f'subm_{i + 1}' argument to the SparseBasicBlock constructor inside the encoder layer loop so that each downsampling stage receives its own unique identifier.
2. Isolated the Configurations: Modified the SparseBasicBlock construction logic to create a shallow copy (conv_cfg.copy()) of the received configuration dictionary before updating it. This breaks Python's mutable dictionary reference chain, ensuring that key assignments from earlier stages do not overwrite or pollute subsequent, downsampled stages.

[CLAassistant]

All committers have signed the CLA.