Fix BlockLinearOperator._getitem slicing for cross-block indices

[robrui]

The modulo check used num_blocks instead of block_size, so slices aligned to individual blocks (e.g. bd[:12, :12] with block_size=12) were incorrectly handled. Cross-block slices (row blocks != col blocks) fell through to block-level indexing instead of the generic path, producing non-square sub-blocks.

Three fixes:

• Use block_size instead of num_blocks for alignment checks
• Detect cross-block slices and fall back to the generic super()._getitem()
• Select blocks from the batch dimension rather than passing block indices as matrix row/col indices