Add optimized YOLOv10n ONNX reference port (721s to ~8s on ET-SoC1)

[ruimouradev]

Summary

Adds a pinned-ONNX YOLOv10n reference port for ET-SoC1 and optimizes it from the
initial working port at 721.396s down to ~7.96s on the board, about 90x,
keeping full FP32 numerical correctness. Model identity is pinned in
artifacts.json (onnx-community/yolov10n, fixed revision and sha256).

Approach

All of the optimization is in src/ref_runtime.c. The graph is memory bound on
ET-SoC1, so the wins come from spreading work across all 16 harts and from the
vector unit, never from lowering precision.

• Conv across harts. Each convolution splits by output channel, every hart
  writing a disjoint cache-line-aligned slice. L1 is minion-local and not
  coherent, so two harts never share a line.
• 1x1 conv on the vector unit. The dominant 1x1 stride-1 path runs eight
  spatial positions at a time with packed FMA, two output channels sharing each
  activation load.
• SiLU fused and vectorized. Sigmoid and Mul fold into one pass using the
  hardware packed exponential and reciprocal, so the intermediate is never
  materialized.
• Serial ops parallelized. Elementwise, Concat, Softmax, MatMul, MaxPool,
  Transpose, Split, Reshape and ReduceMax each ran on a single hart. All of them
  now split across 16 with the same cache-line-aligned publish discipline.
• Depthwise 3x3 on the vector unit, plus an O(1) TopK early reject.

Each step was validated on the board before being kept. Loop-order cache-blocking
of the convs and running them on the tensor engine both regressed on this
memory-bound graph and were dropped.

Results

Board runs on ET-SoC1 (aifoundry2), metric Kernel wait seconds.

build	time
baseline, initial working port	721.396s
this PR, three runs	7.95956s / 7.95906s / 7.95894s
this PR through run_sysemu_model_ports.sh	7.99992s, pass

tools/compare_full.py reports FULL_COMPARE PASS on every run above, 16
checkpoints at atol=5e-5 rtol=1e-4, against a host oracle built from the pinned
ONNX model. FP32 throughout, and the host build is clean under
-Wall -Wextra -Werror.

Benchmark registration

Registered as models.yolov10n_hf_reference so board CI selects and scores it,
with the deterministic input and weight blobs under assets/ as file_loads.
src/leaderboard_full.c is the single translation unit the leaderboard build
wants, compiling the same units and flags the scripts use separately. I built that
ELF on the board with its own toolchain and ran it there as a check: pass,
8.22834s, FULL_COMPARE PASS, same output bytes.

The entry declares src/ as the port's source root so host_full_runner.c and
host_range_runner.c do not trip the coverage check. They are the x86 host oracle
this port validates against, so they cannot be in a board ELF, and the config has
no way to mark a source as host-side. Glad to move them out of src/ instead if
you prefer.

Notes

Outside the port directory this touches only the benchmark_config.json entry, a
.gitignore exception for the committed benchmark blobs (same pattern as yolo
and dncnn), and the model line in the top-level README list. data/*.json and
the generated leaderboard block are left for CI. No submodule changes.