feat: anchor actuation timing on container start (additive)

[rubambiza]

Summary

Adds a container-start timing anchor alongside the existing pod-create anchor, on both the
standalone (no-FMA) path and the FMA analysis output. Everything here is additive — no existing
column, field, or number changes — so nothing that currently consumes the pod-create anchor breaks.

Closes #1836.

Why

The standalone (no-FMA) baseline is measured from Pod creation (includes scheduling + container
startup), while FMA T_actuation is measured from the requester container's start (excludes that
front-matter). Comparing the two paths across a differing start instant is unsound. This surfaces a
consistent container-start number on both paths so a fair, same-anchor comparison is possible —
while keeping the pod-create numbers in place, since we don't know who else consumes them.

See #1836 for the full rationale; this also addresses the review comment.

Changes

• workload/harnesses/nop_functions.py — get_container_start() (elapsed Ready − container
  state.running.startedAt, matches the container by name with a sole-container fallback); a
  container_start field on BenchmarkVllmMetrics, captured and assigned next to pod_start.
• llmdbenchmark/analysis/scripts/nop-analyze_results.py — prints a Container Start(secs) line
  in the vLLM metrics; adds a T_actuation_cstart(s) column (ready − container_start_timestamp) to
  the FMA metrics table (missing anchor renders as --, distinct from a real 0.0).
• llmdbenchmark/analysis/benchmark_report/native_to_br0_1.py — maps container_start through
  the native->benchmark-report import (the hand-mapped converter otherwise dropped it).
• Tests for each (15 new tests), including back-compat: older artifacts without container_start
  default gracefully to 0.0 rather than erroring.

Testing

pytest tests/test_nop_container_start.py tests/test_nop_container_start_metrics.py \
  tests/test_nop_analyze_container_start.py tests/test_nop_analyze_fma_cstart.py \
  tests/test_native_to_br01_container_start.py -q
# 15 passed

Validated end-to-end on live standalone Qwen3-4B runs (container_start ~119s) and on existing FMA
4B result artifacts (container_start_timestamp present on every iteration).

Assisted-By: Claude Code