Re-measure coarse-allocation MILP solve time and its hardware impact

[byeongjee]

Redoes the constant-alloc (coarse-allocation) experiments behind the "Speeding up MILP" subsection, whose numbers were stale.

What was run

• ckpt compile milp for all 13 evaluated benchmarks at 5/10/50 uF, with and without --coarse-allocation (78 compiles, serial, to keep solve times clean) -> baseline_*.csv, coarse_*.csv, summary_*.csv.
• ckpt bench milp on hardware for the 9 benchmark x capacitor pairs where coarse changes the solution, in both modes and with/without device-debug -> hw/.

Results

Geometric mean solve-time reduction: 25.0% at 5 uF, 23.8% at 10 uF, 28.8% at 50 uF (previously reported as 31.6 / 26.2 / 34.2%).

Worst case is now 77.3 s -> 66.5 s on sensor_fusion at 5 uF, not the 5.1 s -> 2.6 s in the current text.

The claim that coarse mode leaves boundary placement and memory allocation unchanged no longer holds: 9 of 39 pairs differ. On hardware, 8 of those 9 run no slower than the per-region solution (up to 2.6% faster); cuckoo_filter at 5 uF is 3.5% slower.

Notes

• The regional re-measurements reproduce the committed results/milp.csv to within 0.04%, so no other measured results are stale.
• rsa at 50 uF alternates between 8 and 9 boundaries across runs in coarse mode (equal-objective solutions); execution time is unaffected.
• milp_solve_time_ms is rounded to whole milliseconds, so the seven pairs solving in 1-5 ms contribute quantization noise to the geomeans.