[codex] Add non-CNN v2 calibration catalog

[JustinLinKK]

Summary

• Adds the opt-in non-CNN/template_v2 catalog path for expanded architecture coverage.
• Adds the shared v2 architecture schema, template catalog, multi-input profiling support, and profile dataset generation.
• Adds stable --hardware-id tagging for profiler/submission outputs and repeatable --results-dir materialization for combined multi-hardware datasets.
• Consolidates root documentation into a single comprehensive README.md and removes the old root Markdown notes.
• Updates .gitignore so generated v2 packs, datasets, profile datasets, smoke outputs, and nrp_noncnn_results_* directories stay out of git.

Dataset Work

• Recreated nrp_calibration_pack_noncnn/ locally as a generated, ignored artifact.
• Generated 10,000 template_v2 base models and 50,000 manifest rows across fp32_ieee,tf32,bf16_amp,fp16_amp,fp8_te_hybrid.
• Generated 10,000 profile dataset specs with 50 train repeats and 50 inference repeats.
• Validated exact family quotas, required v2 manifest fields, schema version, precision counts, and profile spec coverage.
• Full 3090/4090/5090 labels still require running the documented profiler jobs on those hardware classes; FP8 rows are expected to materialize as rejected unsupported rows until GraphModel FP8 support exists.

Tests

• git diff --check
• git ls-files -ci --exclude-standard
• python -m py_compile ... over touched Python files
• python -m unittest tests.test_nrp_calibration_pack tests.test_source_converter -v (89 tests)
• v2 smoke: generated a 15-model template_v2 pack under /tmp, created profile dataset specs, CPU-profiled fp32 with warmup=1, train_repeats=1, infer_repeats=1, materialized labels, and parsed emitted labels to six-target tensors

Notes

• Draft PR opened and updated via GitHub connector because local gh is not installed.