feat: support per-dimension link settings and sync collective dims

[horser1]

related to #32

Summary

• allow link_bw and link_latency to be specified as either scalars or per-dimension arrays
• reuse the same inferred topology dimensions for both network.yml and system.json
• keep the existing default collective implementation as ring
• add docs and a new cluster-config example for distinct intra-/inter-dimension settings

Motivation

ASTRA-Sim supports per-dimension bandwidth/latency when npus_count is multi-dimensional, but LLMServingSim previously only exposed scalar link_bw / link_latency.

While adding array support, I also found that network.yml and system.json were not using exactly the same topology-dimension inference path. In several multi-
instance / PD / DP+EP configs, network.yml could become 2D while system.json still emitted only one collective implementation entry per collective.

This change centralizes topology-dimension inference so both files stay aligned.

Testing

Syntax / parsing

python3 -m py_compile serving/core/config_builder.py serving/__main__.py serving/core/trace_generator.py

python3 - <<'PY'
import json
from pathlib import Path
for path in [
    Path('configs/cluster/dual_node_moe_dp_ep_intra_inter_instance.json'),
    Path('configs/cluster/single_node_single_instance.json'),
]:
    with open(path) as f:
        json.load(f)
    print('OK', path.name)
PY

Result:

- passed

### Config-builder regression

python3 - <<'PY'
import os, sys
from pathlib import Path
repo = Path('/app/LLMServingSim')
os.chdir(repo / 'astra-sim')
sys.path.insert(0, str(repo))
from serving.core.config_builder import build_cluster_config
astra = str(repo / 'astra-sim')
root = repo / 'configs' / 'cluster'
errors = []
for path in sorted(root.glob('*.json')):
    try:
        build_cluster_config(astra, str(path.relative_to(repo)))
        print('OK', path.name)
    except Exception as exc:
        errors.append((path.name, type(exc).__name__, str(exc)))
        print('ERR', path.name, type(exc).__name__, str(exc))
print('SUMMARY', len(errors), 'errors')
PY

Result:

- 14 / 14 cluster configs generated successfully
- 0 errors

### Topology / collective-dimension consistency

python3 - <<'PY'
import json, os, sys, yaml
from pathlib import Path
repo = Path('/app/LLMServingSim')
os.chdir(repo / 'astra-sim')
sys.path.insert(0, str(repo))
from serving.core.config_builder import build_cluster_config
astra = str(repo / 'astra-sim')
root = repo / 'configs' / 'cluster'
keys = [
    'all-reduce-implementation',
    'all-gather-implementation',
    'reduce-scatter-implementation',
    'all-to-all-implementation',
]
for path in sorted(root.glob('*.json')):
    build_cluster_config(astra, str(path.relative_to(repo)))
    with open(repo / 'astra-sim/inputs/network/network.yml') as f:
        net = yaml.safe_load(f)
    with open(repo / 'astra-sim/inputs/system/system.json') as f:
        syscfg = json.load(f)
    ndims = len(net['npus_count'])
    lens = {k: len(syscfg[k]) for k in keys}
    assert all(v == ndims for v in lens.values()), (path.name, ndims, lens)
    print('OK', path.name, ndims, lens)
PY

Result:

- 0 mismatches between len(network.yml::npus_count) and all collective implementation arrays in system.json

### Smoke run

mkdir -p /tmp/llmservingsim-bin
ln -sf "$(command -v python3)" /tmp/llmservingsim-bin/python
PATH="/tmp/llmservingsim-bin:$PATH" \
PYTHONPATH="$(pwd)/astra-sim/extern/graph_frontend:$(pwd)/astra-sim/extern/graph_frontend/chakra" \
python3 -m serving \
  --dtype bfloat16 \
  --cluster-config configs/cluster/single_node_single_instance.json \
  --dataset workloads/example_trace.jsonl \
  --num-reqs 10 \
  --output outputs/pr_smoke_single_10.csv

Result:

- exit code 0
- outputs/pr_smoke_single_10.csv contains 10 data rows
- workload graphs were regenerated under astra-sim/inputs/workload/...
- final summary printed successfully
- observed throughput summary:
    - request throughput: 6.01 req/s
    - average prompt throughput: 72.07 tok/s
    - average generation throughput: 354.94 tok/s

## Notes
- the new example config demonstrates per-dimension intra-/inter-link settings without modifying the existing bundled examples