feat: Tighter FMA timing via DPC log parsing

[rubambiza]

Summary

• Add dpc_log_parser.py module that parses DPC controller logs to extract
  httpCallStartTime/k8sCallStartTime fields at microsecond precision
• Integrate parser into benchmark_fma() to override Kube-timestamp upper
  bounds with tighter DPC-derived values when available
• Graceful fallback: if DPC log data is unavailable (pre-PR Log call durations on launcher/instance/K8s actuation calls llm-d-incubation/llm-d-fast-model-actuation#522 DPC, low
  verbosity, requester crash before relay), Kube upper bounds stand unchanged
• Streams log files line-by-line to handle large DPC logs without memory pressure
• Propagate dpc_timing_available through native_to_br0_1 conversion so the
  analysis report attributes each row to its true source
• Adds "Source" column (DPC/Kube) to analysis output

Closes #1465

Timing improvement (validated on cluster)

Measured on a 10-iteration nop run (scale 0->1->0 per iteration). The DPC log
parser now yields sub-second hot-start timing, confirming that the previous
8-10s figures were dominated by the kubelet readiness-probe interval rather than
actual wake latency.

Path	Before (Kube upper bound)	After (DPC log)	Source of inflation
Hot	8s (also 3s when probe lands sooner)	0.53-0.59s (t_wake)	kubelet readiness probe delay (~10s period)
Warm	53-62s (T_actuation)	~50.5s (t_instance_create)	requester scheduling + probe delay
Cold	n/a this run	--	no cold-launcher iterations occurred in the validation run

An independent cross-check from the vLLM launcher log on the same run reported a
wake transition of ~0.51s, consistent with the DPC-derived t_wake.

Measurement conditions

• Model: meta-llama/Llama-3.1-8B-Instruct
• FMA launcher image: ghcr.io/llm-d-incubation/llm-d-fast-model-actuation/launcher:v0.6.0-alpha.13 (includes PR Unexpected EOF from inference-perf-analyze_results.sh #522 timing fields)
• vLLM: 0.20.2, sleep mode enabled, gpu_memory_utilization=0.95, prefix caching off
• Cluster: OpenShift on IBM Spectrum Scale storage, NVIDIA GPUs
• Run: 10 iterations; hit rates this run were 30% hot / 70% warm / 0% cold

Numbers are representative of a single validation run; hot-start hit rate (and
thus how often the sub-second path is exercised) depends on workload pattern and
launcher capacity configuration.

Test plan

• Unit tests for klog line parsing (hot/warm/cold paths)
• Edge cases: retry, unready relay, malformed timestamps, multi-requester
• File-based parsing (directory scanning, missing dir, partial logs)
• Large-log detection: indicator message located far into the file (past the initial read window)
• Integration run on cluster with DPC v0.6.0-alpha.13+ (PR Unexpected EOF from inference-perf-analyze_results.sh #522 merged)
• Verify timing values match manual log inspection (vLLM launcher log cross-check)

Related

• Depends on: Log call durations on launcher/instance/K8s actuation calls llm-d-incubation/llm-d-fast-model-actuation#522 (merged)
• Builds on: feat: FMA actuation path classification, hit rates, and per-path timing #1429 (upper-bound timing + path classification)

Assisted-By: Claude Code

[rubambiza]

The force push (5d06f1d) was a fresh rebase on main.