perf: Implement HTTP client connection pooling in LLM clients

[AlexMikhalev]

Issue Description

New HTTP clients are created for each LLM request instead of reusing connections, causing significant overhead.

Location

crates/terraphim_service/src/llm.rs (lines 285, 540, 569)

Current Code

```rust
// Creating new client for each request
let client = reqwest::Client::new();
// ... use client once ...
```

Impact

• CRITICAL priority - Performance
• Creates new connection pool for every client initialization
• 30-50% connection overhead for repeated requests
• Increased memory usage

Recommended Fix

```rust
use once_cell::sync::Lazy;
use reqwest::Client;

static HTTP_CLIENT: Lazy = Lazy::new(|| {
Client::builder()
.pool_max_idle_per_host(10)
.pool_idle_timeout(Duration::from_secs(90))
.timeout(Duration::from_secs(120))
.build()
.unwrap()
});

// Use: &*HTTP_CLIENT instead of creating new clients
```

Expected Improvement

• 30-50% reduction in connection overhead
• Reduced memory allocations
• Faster repeated requests

References

• Identified in performance analysis for PR fix(zipsign): update zipsign API to v0.2 #429
• Related to LLM client architecture