feat(profiling): expose common nsys flags - #412

Summary

• expose trace_domain, cuda_graph_trace_mode, and sample_mode in the profiling recipe
• apply explicit overrides consistently to TRT-LLM and non-TRT-LLM nsys commands
• preserve every existing backend-specific default when the new keys are omitted
• document the settings and cover both time- and iteration-based paths

Testing

• uv run ruff check src/srtctl/core/schema.py tests/test_profiling.py
• uv run pytest tests/test_profiling.py -v --tb=short (34 passed)
• git diff --check

A full-suite run reached unrelated existing failures: test_apply_mock_emits_submission_json_and_spawns_worker has no effective CPU allocation on this host, and the SA-Bench script tests cannot locate profiling.sh when sourced from the test harness.