[TOGSim] Report off-chip DRAM energy in the end-of-run stats

[YWHyuk]

Adds an off-chip DRAM energy/power figure to the end-of-run TOGSim stats.

E = n_ACT x row_activation_pj + bits x sum(transfer_pj_per_bit)

Where the counters come from

n_ACT — Ramulator2 has no ACT command counter, so activations are derived from the per-controller row stats that already exist: a row hit needs no ACT, a miss (closed bank) and a conflict (another row open) each need one, so n_ACT = row_misses + row_conflicts, summed over channels.

bits — served transactions times the DRAM model's own get_tx_bytes() (internal_prefetch_size * channel_width / 8, 32 B for the HBM2 config), times 8. A DRAM column command moves one full burst in bytes, never a partial one, so the byte granularity follows the ramulator config instead of a hardcoded size and is only converted to bits at the last step. The report prints the transaction count and burst size so the conversion is auditable.

Output

A separate section at the end of a run, aggregated over all channels. Periodic interval logs are untouched.

=== Energy statistics ===
[Energy] Energy cost table "HBM2" loaded from "configs/energy_tables/hbm2.yml"
[Energy] OffChip DRAM channels 0..15 combined | 11.085 W avg power, 3.358 mJ over 302.879 us | 555.087 uJ activation, 2.802 mJ transfer
[Energy] OffChip DRAM channels 0..15 combined | 610657 activations x 909.00 pJ | 3145728 transactions x 32 B x 3.48 pJ/bit (dram 1.51 + io 1.17 + phy 0.80)

Configuration

energy_cost_table_path: ../configs/energy_tables/hbm2.yml

Omit the key and no energy section is printed, so every existing config behaves exactly as before. The constants live in a swappable table (configs/energy_tables/hbm2.yml) with the per-bit terms kept separate so the report can show the breakdown. Activation energy needs row state, so the report requires dram_type: ramulator2; with simple it prints a one-line note instead.

Verification

• 2048-cube matmul: the reported 610657 activations match the sum of row_misses and row_conflicts across the 16 per-channel ramulator dumps exactly. row_hits + row_misses + row_conflicts equals the transaction count, so both terms are counted over the same event population.
• Arithmetic checks out by hand: 610657 x 909 pJ = 555.087 uJ; 3145728 x 32 B x 8 x 3.48 pJ = 2.802 mJ; 284706 cycles at 940 MHz = 302.879 us, giving 11.085 W.
• Regression: a config without energy_cost_table_path produces byte-identical stats output to before, and tests/ops/elementwise/test_add.py and tests/ops/gemm/test_matmul.py pass.

Scope

Dynamic, activity-proportional energy only. Refresh and background/standby power are not modelled, so the figure is a floor on real device power. Effective energy works out to 3.93 pJ/bit at 512-cube and 4.17 pJ/bit at 2048-cube, and 16 channels is about two HBM2 stacks, so roughly 5.5 W per stack at 69% bandwidth utilisation.

Depends on

PSAL-POSTECH/ramulator2#1, which adds the row_activations() accessor. The submodule pointer in this PR references that branch and needs re-pointing at the merged commit once it lands.