perf: add criterion benchmarks for hot paths

## Context

Only markdown benchmark exists. Need baselines for: token estimation, context compaction, skill matching.

## Proposed benchmarks

1. `zeph-memory/benches/token_estimation.rs`
2. `zeph-core/benches/context_building.rs`
3. `zeph-skills/benches/matcher.rs`

## Affected crates

zeph-core, zeph-memory, zeph-skills

## Ref

`.local/plan/performance-audit.md` — Issue 7.1