…ase (flagos-ai#1145)

### PR Category
CICD

### PR Types
New Features

### PR Description

Add a benchmark functional test framework to the CI/CD pipeline,
enabling automated detection of training performance regressions by
comparing runtime metrics against baseline gold values.

- Introduce dedicated benchmark workflow
(`functional_tests_benchmark.yml`) with artifact download retry logic,
environment auto-setup, benchmark log parsing, metrics upload to
backend, and failure log archiving
- Add `test_benchmark_equal` in `check_results.py` supporting two
threshold types (`upper_bound` for elapsed time, `lower_bound` for
throughput) with configurable tolerance and automatic warmup iteration
skipping (first 5 iterations)
- Add `parse_benchmark_output.py` to extract metrics from training logs
into structured JSON and upload to the metrics backend for tracking
- Register `benchmark` as a new task type in the test runner
(`run_functional_tests.sh`), platform config loader
(`load_platform_config.sh`), and platform definition (`cuda.yaml`)
- Add Qwen3 tp2_pp2 as the first benchmark test case with gold values
for elapsed time per iteration (ms) and throughput per GPU (TFLOP/s/GPU)
at 10% tolerance
- Integrate benchmark tests into `all_tests_common.yml` and re-enable
inference/serve test checks in the completion gate

---------

Co-authored-by: zihugithub <fbye@baai.ac.cn>