perf: tune skill embedding concurrency limit

[bug-ops]

Problem

Skill matcher uses buffer_unordered(50) with no tuning or timeout, potentially spawning 50 concurrent HTTP requests.

File: crates/zeph-skills/src/matcher.rs lines 20-37

Current code:

stream::iter(skills.iter().enumerate())
    .map(|(i, skill)| { embed_fn(&skill.description) })
    .buffer_unordered(50)  // Unbounded for large skill sets
    .filter_map(|x| async { x })
    .collect()
    .await;

Impact

• Memory: 175KB peak with 50 skills × 3.5KB embeddings
• Network: May hit API rate limits
• No timeout: Slow embeddings block entire batch

Solution

1. Reduce concurrency to 20 (optimal for network I/O)
2. Add per-embedding timeout:

.map(|(i, skill)| {
    let fut = timeout(Duration::from_secs(5), embed_fn(&skill.description));
    // handle timeout
})
.buffer_unordered(20)

Priority: P1
Effort: Small (1 hour)
Related to #391