Improve AMS performance & benchmarking: semaphore fast-path, checksum, bench stats

[zebastian]

• Inline semaphore hot path (_semTbl, _semGetSem, _semSync) and split _semSync into inline fast-path + noinline slow-path to avoid function call overhead on the common (seq-match) case
• Optimize computeAmsChecksum to process 2 bytes per iteration instead of one-at-a-time with branch
• Defer keyBuffer initialization in recoverMsgContent to the branch that actually uses it
• Add cell census log event in amsd to enable event-driven startup instead of fixed sleep (speedup benchmark)
• Enhance amsbenchr with per-message inter-arrival timing, percentile stats (p50/p95/p99), jitter, out-of-order detection, and min/max/avg message size reporting
• Add variable message size support to amsbenchs (min/max range)
• Extract wait_for_log_event helper (bench_helpers) and refactor ionstart scripts to wait for census log event instead of fixed 9s sleep
• Update dotest to use MESSAGE_SIZE_MIN/MAX variables

Note: some changes were developed with the help of AI (claude code), especially the amsbenchr.* changes. I double checked execution time: instructions and per function cost in kcachegrind. ETA enhancement of possible throughput (or reduction in cpu time) for continious flow of messages around 20-30%.

Benchmark results:
Received 1000 messages, a total of 1000000 bytes,in 0.310885 seconds.
3216.624 messages per second.
24.541 Mbps.
Message size: min=1000 max=1000 avg=1000 bytes.
Out-of-order: 0 of 1000 messages (0.00%).
Inter-arrival (us): min=80 max=13352 avg=311 p50=282 p95=356 p99=392 Mean jitter (us): 55.0