Add the load cases A and B

[tamnd]

Part of #1.

What changes

• src/load.rs has the steps of spec/20 section 20.3.1. Case A runs sync and drop_caches, checks with fincore that the source has 0 resident pages, times cat <source> > /dev/null as R, drops the caches and checks again. Case B reads the source twice with cat, and fincore must report at least 99 percent of the pages resident. Below 99 percent, the run does the steps of case A and records case A with a note.
• A few pages can stay resident after one drop of the caches. The first try on server3 left 20 pages of a file that had just been downloaded. The check therefore drops the caches up to 5 times until fincore reports 0, and the result records how many drops each check needed. If pages stay after 5 drops, the run fails, because then another process holds the file or the file is on tmpfs.
• rupg-bench load --case a|b --source FILE (--name N | --attach PATH | --unit UNIT) [--du PATH] [--json] -- COMMAND runs the steps and then times the load command in the cgroup runner of Add the cgroup v2 runner and the measure command #11, so the resources of spec/20 section 20.3.4 are measured over the load. The result has L, R, L / R, every fincore line and the case that was asked for.
• src/cgroup.rs gets run, which starts a command in the cgroup (or outside it, for a client) and measures the interval. measure and load share it.
• The README describes rupg-bench load.

Test

All runs are on server3 (8 cores, Linux 6.8.0-106-generic) on 7 October 2026, with DuckDB 1.5.6 from machines/install/duckdb.sh. The source is the first file of the partitioned ClickBench data, hits_0.parquet (122446530 bytes, sha256 fa134fe101e68324e0de851146fda69624f5cbb707d387141d1c2a88a219a16d), from https://datasets.clickhouse.com/hits_compatible/athena_partitioned/hits_0.parquet. These are smoke runs that show the steps work. They are not baselines: the machine is shared, its load average was near 75, and the file is a 1 percent sample.

Case A:

rupg-bench load --case a --source /root/work/data/hits_0.parquet --name duckdb-load --du /root/work/m/h.db -- \
  /root/work/rupg-bench-opt/bin/duckdb /root/work/m/h.db -c "CREATE TABLE hits AS SELECT * FROM read_parquet('/root/work/data/hits_0.parquet'); CHECKPOINT;"

case            A (asked for A)
fincore         0.00 percent resident, case A, before the read: 0 0 122446530 /root/work/data/hits_0.parquet
fincore         0.00 percent resident, case A, before the load: 0 0 122446530 /root/work/data/hits_0.parquet
drops           1, 1 (drops of the caches before each check)
R (cat)         6.946 s
L (load)        73.977 s
L / R           10.650
memory.peak     1634.73 MiB (1714135040 bytes) (since the cgroup was created)
io read         141.44 MiB (148312064 bytes)
io write        180.66 MiB (189440000 bytes)
du /root/work/m/h.db  180.51 MiB (189280256 bytes)

Case B, with --json. fincore found all 29895 pages resident after the two reads, and the load read only 13451264 bytes from the device, so the source came from the page cache:

"run": { "wall_s": 121.894304608, "rbytes": 13451264, "wbytes": 188915712, "memory_peak_bytes": 1576206336, ... },
"load_case": { "case_wanted": "B", "case": "B", "note": null, "read_s": null,
  "fincore": [ { "resident_pages": 29895, "size_bytes": 122446530, "resident_share": 1, "step": "case B, after two reads" } ] }

The case B load took longer than the case A load because other jobs on the machine took the CPU. This is why these numbers are not baselines.

srvtest passed on server3: fmt, clippy with -D warnings, and 23 tests, with new tests for the fincore parser, the resident share, L / R and the case names.