ZTS: Fix L2ARC test reliability

Disable depth cap (L2ARC_EXT_HEADROOM_PCT=0) in DWPD and multidev tests
that rely on predictable marker advancement during fill and measurement.

Rework multidev_throughput to verify sustained throughput across three
consecutive windows instead of asserting an absolute rate. Use larger
cache devices (8GB) to avoid frequent global marker resets
(smallest_capacity/8), fill ARC before attaching caches to provide a
stable evictable buffer pool, and lower write_max to 8MB/s to avoid
exhausting data within the measurement period.

Use destroy_pool (log_must_busy) instead of log_must zpool destroy
to handle transient EBUSY during teardown.

Remove l2arc_multidev_throughput_pos from the expected-fail list in
zts-report.py.in.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <alexander.motin@TrueNAS.com>
Signed-off-by: Ameer Hamza <ahamza@ixsystems.com>
Closes #18321

tests/test-runner/bin/zts-report.py.in

@@ -245,7 +245,6 @@ maybe = {
     'history/history_010_pos': ['SKIP', exec_reason],
     'io/mmap': ['SKIP', fio_reason],
     'l2arc/l2arc_l2miss_pos': ['FAIL', known_reason],
-    'l2arc/l2arc_multidev_throughput_pos': ['FAIL', 18272],
     'l2arc/persist_l2arc_005_pos': ['FAIL', known_reason],
     'largest_pool/largest_pool_001_pos': ['FAIL', known_reason],
     'mmap/mmap_sync_001_pos': ['FAIL', known_reason],