Count the cache-line transactions needed for a group of memory accesses. Each lane requests access_width_bytes consecutive bytes from its starting address. For lane $i$, $a_i$ is the start address, $W$ is access_width_bytes, $L$ is cache_line_bytes, and $s_i$ and $e_i$ are its first and last touched line IDs. Floor rounds down.

Each distinct touched line costs one transaction. Let $U$ be the number of distinct requested byte addresses across all lanes and $C$ be the number of touched lines:

Overlapping requests count shared bytes only once. Return a dictionary containing line_ids as a sorted one-dimensional NumPy int64 array of length C, transaction_count as a Python integer, and useful_byte_fraction as a Python float.