Describe the changes
This PR:

• Fixes a flaw in the CUDA C++ that meant batches of more than 8192 chunks would segfault
• Fixes another flaw in the CUDA C++ that would cause corrupt data if you overprovisioned chunks (requested more chunks to be packed than you have blocks * threads per block)
• Adds logic to control the blocks and threads per block of the packing kernel
• Adds logic to automatically best-guess the blocks and thread per blocks
  NB: testing has shown that we get the best performance by trying to match the underlying hardware - this means blocks = SM count, and threads per block = physical CUDA code count per SM. There is likely room for improvement here with more indepth testing (using nsight etc).

Checklist

• [] Tests have been added/updated for the changes.
• Documentation has been updated for the changes (if applicable).
• The code follows Rust's style guidelines.