Configurable sampling per test (1/N or percentage)

## Problem
Some tests should run on a small fraction of matched traffic to reduce overhead/impact.

## Proposal
Add per-test sampling:
- fixed percentage or 1/N
- stable hashing to keep flow-consistent sampling

## Alternatives considered
External sampling before OpenPenny.

## Additional context
Useful when expanding a rule to large prefixes.