Add syslog adaptor benchmarks

[ruslandoga]

Please ignore this PR for now :)

---

Adds benchmark-only syslog profiling scripts under test/profiling.

The benchmarks compare current batcher routing, no-batcher per-message sends, and preformatted batch sends using a local RFC6587 octet-counting TCP sink.

The Broadway benchmark runs through real Broadway processors/batchers, BufferProducer, a local NimblePool TCP sender, and fake cache entries so it can run with mix run --no-start.

Validated with mix format and small smoke runs for both benchmark scripts.

[ruslandoga]

Benchmark smoke results from this workspace. These used tiny inputs to validate the scripts and the RFC6587 sink wiring, so treat the relative ordering as a sanity check only, not as production throughput data.

Direct formatter + TCP send:

Command:
SYSLOG_BENCH_EVENTS=20 SYSLOG_BENCH_MESSAGE_BYTES=200 SYSLOG_BENCH_BATCH_SIZE=5 SYSLOG_BENCH_TIME=0.1 SYSLOG_BENCH_WARMUP=0 mix run --no-start test/profiling/syslog_send_bench.exs

##### With input 200 byte message #####
Name                                                                          ips        average  deviation         median         99th %
handle_message routes, handle_batch formats+sends batch                    957.01        1.04 ms   +/-260.07%        0.74 ms       27.09 ms
no batcher: handle_message formats+sends each message                      466.68        2.14 ms    +/-64.21%        1.83 ms       10.36 ms
preformat in handle_message, handle_batch sends preformatted frames        275.50        3.63 ms    +/-98.59%        2.17 ms       15.66 ms

Comparison:
handle_message routes, handle_batch formats+sends batch                    957.01
no batcher: handle_message formats+sends each message                      466.68 - 2.05x slower +1.10 ms
preformat in handle_message, handle_batch sends preformatted frames        275.50 - 3.47x slower +2.58 ms

Broadway end-to-end:

Command:
SYSLOG_BENCH_EVENTS=20 SYSLOG_BENCH_MESSAGE_BYTES=200 SYSLOG_BENCH_BATCH_SIZE=5 SYSLOG_BENCH_CONCURRENCY=1 SYSLOG_BENCH_PUSH_CHUNK=10 SYSLOG_BENCH_TIME=0.1 SYSLOG_BENCH_WARMUP=0 mix run --no-start test/profiling/syslog_broadway_bench.exs

##### With input 200 byte message #####
Name                                                                                  ips        average  deviation         median         99th %
broadway: no batcher, handle_message formats+sends each message                     80.99       12.35 ms     +/-3.29%       12.17 ms       12.96 ms
broadway: handle_message routes, handle_batch formats+sends batch                   74.04       13.51 ms     +/-6.17%       13.38 ms       15.01 ms
broadway: handle_message preformats, handle_batch sends preformatted frames          6.86      145.79 ms     +/-0.00%      145.79 ms      145.79 ms

Comparison:
broadway: no batcher, handle_message formats+sends each message                     80.99
broadway: handle_message routes, handle_batch formats+sends batch                   74.04 - 1.09x slower +1.16 ms
broadway: handle_message preformats, handle_batch sends preformatted frames          6.86 - 11.81x slower +133.44 ms

Interpretation:

The direct send benchmark is mostly isolating formatter cost plus the number of TCP sends. With this tiny input, batching wins clearly because it does fewer :gen_tcp.send/2 calls and the local sink waits until all expected octet-counted frames are parsed.

The Broadway smoke run includes pipeline startup, BufferProducer polling, processor/batcher handoff, NimblePool checkout, TCP send, and receiver parsing. At this very small event count, fixed Broadway overhead dominates, so the no-batcher path can look slightly faster than the batched path. The preformat variant is much slower here because it pays formatting cost in handle_message/3 and still goes through batcher coordination before sending.

For useful throughput comparison, run the same scripts with larger inputs, for example:

SYSLOG_BENCH_MESSAGE_BYTES=200,2000,50000 \
SYSLOG_BENCH_CONCURRENCY=1,5 \
SYSLOG_BENCH_BATCH_SIZE=50 \
SYSLOG_BENCH_EVENTS=5000 \
SYSLOG_BENCH_PUSH_CHUNK=1000 \
mix run --no-start test/profiling/syslog_broadway_bench.exs

[Ziinc]

instead of copying the pipeline, can we make the pipeline parameterized such that we can pass opts to it, see BigQuery.Pipeline for how the args are passed for benchmarking

[Ziinc]

can be condensed.