fix(build): cap ONNX Runtime build parallelism to avoid runner OOM

[mc-nv]

What does the PR do?

Caps the ONNX Runtime build parallelism in the generated Dockerfile so fully-uncached
builds no longer OOM the GitLab runner hosts (the "nested ONNX Runtime build" called out in
TRI-1550).

Previously gen_ort_dockerfile.py passed a bare --parallel to ORT's build.sh, which
expands to multiprocessing.cpu_count(). Combined with per-nvcc multi-arch compilation
(7 CUDA archs), that exhausted memory on swap-less builders and the OOM killer terminated
runner/system daemons instead of the build.

The parallelism is now computed in Python at Dockerfile-generation time and baked into the
build step as literals — e.g. --parallel 16 --nvcc_threads 2:

• usable_cpu_count() honors CPU affinity / cgroup pinning (Linux), with a cpu_count()
  fallback.
• available_memory_gb() reads MemAvailable from /proc/meminfo.
• build_parallelism() budgets ~2 GB of RAM per concurrent compiler slot and bounds the
  product parallel_jobs * nvcc_threads by that budget.
• The assumed budget is capped at MAX_BUILD_MEMORY_GB (64 GB): never plan for more than
  that, nor for more than is actually available; assume the cap when memory is unknown.

Baking literals keeps the build stage free of shell logic, so it is portable across the
Debian (apt) and RHEL (dnf) base-image paths.

Related PRs:

triton-inference-server/server#8873

Where should the reviewer start?

tools/gen_ort_dockerfile.py — the new usable_cpu_count() / available_memory_gb() /
build_parallelism() helpers, the MAX_BUILD_MEMORY_GB constant, and the ./build.sh
RUN step (bare --parallel removed from COMMON_BUILD_ARGS, now
--parallel <jobs> --nvcc_threads 2).

Test plan:

• Verified the generated Dockerfile RUN line renders as clean literals (no shell logic) and
  Python parses.
• Validated the parallelism math across host profiles:
  • 128 c / 512 GB → --parallel 16 --nvcc_threads 2 (budget capped at 64 GB, ~64 GB peak)
  • 32 c / 48 GB → --parallel 12 --nvcc_threads 2
  • 16 c / 32 GB → --parallel 8 --nvcc_threads 2
  • 8 c / 4 GB → --parallel 1 --nvcc_threads 2
  • memory unknown → assumes 64 GB, bounded by cores
• Validate on the affected builder / arm-builder jobs that a fully uncached build no
  longer OOMs the runner.

Caveats:

• The 64 GB budget and ~2 GB/slot figure are named constants at the top of the file, easy
  to retune.
• Values are computed on the host running gen_ort_dockerfile.py (same runner that then
  runs docker build); a --memory-limited build container smaller than host RAM is not
  reflected. Same model as the server/build.py change.

Background

GitLab build jobs were failing with logs abruptly stopping because runner hosts OOM'd
during fully uncached builds — unbounded parallelism in the nested ONNX Runtime build on
swap-less builders.

Related Issues:

• Resolves: TRI-1550