Expose per-grid dilation amounts in the Python dilate API

[blackencino]

Summary

The Python Grid.dilated_grid() and GridBatch.dilated_grid() methods currently accept a single int dilation value that is applied uniformly to every grid in the batch. The underlying C++ implementation (GridBatchImpl::dilate) already supports a std::vector<int64_t> overload that allows specifying a different dilation amount per grid in the batch.

Motivation

Per-grid dilation amounts would enable use cases where different grids in a batch need different dilation radii -- for example, adaptive resolution workflows or per-object morphological operations in a batched pipeline.

Proposed Change

Extend the Python API so that dilated_grid accepts either:

• A single int (current behavior, applied to all grids), or
• A list/sequence of int with one entry per grid in the batch.

This would require:

1. Adding or overloading the pybind11 binding in GridBatchBinding.cpp to accept std::vector<int64_t>.
2. Updating GridBatch.dilated_grid() in grid_batch.py and Grid.dilated_grid() in grid.py to accept int | list[int].
3. Adding Python-level tests for the per-grid dilation path (including mixed zero/non-zero dilation amounts).

Additional Context

The C++ op dilateGrid in BuildDilatedGrid.cu already handles per-grid dilation correctly, including the edge case where individual dilation amounts are zero.