perf(e2e): reduce test case count to optimize CI execution time (#797)

Reduce test cases for the three slowest tests in dynamic-config profile:
- domain-classify: 280 → 56 cases (4 per category across 14 categories)
- jailbreak-detection: 100 → 25 cases (every 4th case)
- pii-detection: 100 → 25 cases (every 4th case)

Total reduction: 480 → 106 test cases (78% reduction)

Rationale:
Testing revealed the bottleneck is LLM inference time (~5-24s per request),
not test framework overhead. Concurrent requests cause HTTP 500 errors as
the backend cannot handle parallel load. Representative sampling maintains
test coverage while significantly reducing CI execution time.

Expected CI time savings:
- domain-classify: 38m 38s → ~8m
- jailbreak-detection: 39m 52s → ~10m
- pii-detection: 14m 40s → ~3m 40s
- Total: ~93 minutes → ~22 minutes (~71 minute savings)
- Overall CI: 114 minutes → ~43 minutes (62% faster)

Issue: #796

Signed-off-by: Yossi Ovadia <[email protected]>

Author: yossiovadia