feat(trace): add agk trace diff to compare two runs

[kunalkushwaha]

What

Adds agk trace diff — compare two trace runs to answer "did my change help?". This is feature B5 from the FEATURES.md roadmap, and it pairs naturally with agk run's trace summary (#13) and the per-model cost work (#15).

Behavior

Diffs two runs across duration, spans, LLM calls, tokens, and estimated cost, showing the delta with an arrow + percentage. For duration/tokens/cost (lower-is-better), improvements are green and regressions red.

📊 Trace Diff
  A (baseline): run-200-base
  B (new):      run-201-new
────────────────────────────────────────────────────────────────
METRIC     A (baseline)  B (new)  Δ
Duration   3.20s         2.10s    -1.10s ▼ -34%
Spans      8             6        -2 ▼ -25%
LLM Calls  3             2        -1 ▼ -33%
Tokens     2000          1200     -800 ▼ -40%
Est. Cost  $0.0200       $0.0120  -$0.0080 ▼ -40%

Run selection

Form	Compares
agk trace diff	the two most recent runs
agk trace diff <a>	<a> (baseline) vs the latest run
agk trace diff <a> <b>	the two explicit runs

Testing

• go build, go vet, go test ./..., gofmt all green.
• Unit tests (the cmd package's first) for delta direction, the metric set, formatting, and explicit run resolution.
• Verified end-to-end with synthetic manifests: explicit pair, zero-arg (picks two most recent), and the "not enough runs" error path.

Independent branch off main (alongside #13–#16). Reuses the existing readManifest/TraceRun helpers in cmd/trace.go.

🤖 Generated with Claude Code