Nccl translator

[Amos-Rai-KEYS]

Add NCCL topology translator

Summary

This branch adds a new NCCL translator that parses NCCL-generated topology XML into the infragraph device/infrastructure model, alongside annotations capturing per-GPU and per-NIC metadata. It also generalizes the translator framework to support multiple tools and adds a --device_name option.

What's new

NCCL translator (src/infragraph/translators/nccl_translator.py — 604 lines)

• NcclParser — parses NCCL XML topology into a Device/Infrastructure graph: CPUs and CPU fabric links, PCI bridge hierarchy, GPUs/XPUs, NICs, and NVLink/NVSwitch peer connections.
• NcclHelper.generate_nccl_topology() — generates a fresh topology XML at runtime by loading libnccl.so/libcudart.so and calling ncclCommInitAll (used when no --input file is supplied).
• get_annotations() — produces an InfraGraphService annotated with per-GPU metadata (rank, busid, device model, CPU affinity) and per-NIC metadata.
• run_nccl_parser() — writes both the translated device file and an annotated_infragraph.json (infrastructure + annotations).

Translator framework changes

• translator_handler.py — adds nccl to supported translators and threads a device_name argument through; requires --device_name for nccl.
• __main__.py — new --device_name CLI option; default output renamed dev.yaml → device.yaml.

Tests & notebooks

• test_nccl_translator.py — validates component counts, CPU/NVSwitch parsing, edges/links.
• dgx_a100_nccl_topo.xml — DGX A100 mock fixture.
• nccl_translate.py — example notebook.