feat: argus gcwhy — narrate the worst recent GC pause

[rlaope]

Closes #151.

Summary

argus gcwhy <gc-log-file> [--last=5m] picks the single worst pause in the lookback window and emits up to three plain-English "why" bullets plus a related-counters block.

Sample output:

┌─ GC Why ────────────────────────────────────────────────────┐
│  Worst pause: 912 ms at t=14:02:11  (Mixed / G1 Evacuation) │
│                                                             │
│  Why this happened:                                         │
│   • Allocation rate surged roughly 6.8x the recent baseline │
│     in the events leading up to this pause — a burst of     │
│     allocation pressure.                                    │
│   • Heap was 94% occupied at pause start — very little      │
│     headroom left for new allocation.                       │
│   • Pause was 4.2x the recent average (912 ms vs 220 ms) —  │
│     this is an outlier.                                     │
│                                                             │
│  Related counters:                                          │
│    pause-ms:              912                               │
│    type:                  Mixed                             │
│    cause:                 G1 Evacuation Pause               │
│    heap-used-pct:         94.0                              │
│    prior-avg-pause-ms:    220                               │
└─────────────────────────────────────────────────────────────┘

Rule engine (ordered)

1. Explicit System.gc() trigger
2. G1 humongous allocation
3. Metaspace pressure (Metadata GC Threshold)
4. Full GC fallback (concurrent-mode / evacuation failure)
5. Allocation burst vs recent baseline (heap-before jump)
6. High heap occupancy at pause start (≥ 90%)
7. Outlier pause vs recent average (≥ 3× prior avg)
8. Fallback neutral one-liner when nothing fires

Emits the top three triggered bullets (ordered). Always at least one bullet.

Scope

• Log-based: reuses GcLogParser, filters events to --last=Nm window
• --format=json for automation
• --last accepts Ns, Nm, Nh or a bare number in seconds; default 5m
• i18n (en, ko, ja, zh), shell completions (bash, zsh, fish, ps1)

Non-goals in this PR

• Live PID / target mode — deferred to a follow-up; requires per-event GC histograms over /prometheus that are not yet exposed.

References (design lineage)

• Censum-style reasoning ("why" not just "what")
• JDK Mission Control GC log narrative

Test plan

• ./gradlew :argus-cli:test --tests io.argus.cli.gcwhy.* — passes
• ./gradlew :argus-cli:test full suite — passes
• Unit tests cover: System.gc, humongous, allocation burst, high occupancy, Full GC fallback, empty events, window filter picking the right in-window event