[Wrapper] Expose row activation counts for energy accounting

[YWHyuk]

Groundwork for the off-chip DRAM energy report in PyTorchSim (companion PR on PSAL-POSTECH/PyTorchSim). TOGSim needs a per-memory-system activation count to price row activations; nothing here changes simulation behaviour.

What

• Ramulator2::row_activations() — Ramulator2 has no ACT command counter, so this derives activations from the per-controller row stats that already exist: a row hit needs no ACT, while a miss (closed bank) and a conflict (another row open) each need one. It walks the memory system's collect_stats() tree and sums row_misses + row_conflicts, so the derivation stays out of the controller sources and no stat registration changes.
• Transaction counters widened from int to uint64_t. They overflow at roughly 2 billion transactions, which a long run reaches.

Caveats, documented in the header

• Refresh-driven activations are not counted.
• The row stats are updated once per request at scheduling time, so this is the standard approximation of issued ACT commands rather than an exact command count.

Verification

On a 2048-cube matmul through TOGSim, row_activations() summed over 16 channels returns 610657, which matches the sum of row_misses and row_conflicts across the 16 per-channel YAML stat dumps exactly.