rapl_include_dram=True reads DRAM domains but never adds them to the energy total

On Linux, rapl_include_dram=True selects the DRAM RAPL domains and reads them every cycle, but their energy never reaches the reported total. The measurement is silently identical to rapl_include_dram=False, while the log claims DRAM is included.

Reproduction

On a machine exposing a top-level dram domain alongside package-0:

a = EmissionsTracker(rapl_include_dram=False)
b = EmissionsTracker(rapl_include_dram=True)

Both report the same CPU energy, even though b logs RAPL - Including 1 DRAM domain(s) for complete hardware power measurement (CPU+DRAM).

Root cause

Two naming conventions that were never tied together.

• codecarbon/core/cpu.py:708 — domain selection correctly appends DRAM domains when rapl_include_dram is set.
• codecarbon/core/cpu.py:772 — _create_rapl_files only assigns the Power-Gadget-compatible display name Processor Energy Delta_N(kWh) to domains whose name contains package or psys. A DRAM domain falls into the else and keeps its raw sysfs name dram.
• codecarbon/external/hardware.py:390 — _get_energy_from_cpus sums only keys matching ^Processor Energy Delta_\d, so the dram entry is dropped. Same for power: codecarbon/core/cpu.py:849 only derives a Processor Power key from names containing Energy.

Windows EMI does not have this bug: codecarbon/core/windows_emi.py:581 names every selected channel Processor Energy Delta_N(kWh), DRAM included. So emi_include_dram=True works and rapl_include_dram=True does not.

Expected vs actual

Expected: with rapl_include_dram=True, reported CPU energy is package + DRAM (DRAM is typically 10-20% of package power on memory-heavy workloads).

Actual: DRAM energy is read, converted, logged, and discarded. Reported energy equals the package-only value.

Note this is a measurement change for anyone already setting the flag — their reported figures will (correctly) increase.

Start with the RAPL flow in codecarbon/core/cpu.py at lines 708, 772, and 849, then trace aggregation in codecarbon/external/hardware.py:390 and compare the working naming path in codecarbon/core/windows_emi.py:581. Reproduce the package-only versus DRAM-enabled measurements and verify that DRAM contributes to the reported energy and power totals when rapl_include_dram=True.