Small optimizations

[KA-ROM]

Summary

Six small, self-contained optimizations on the engine_* hot path, each removing a per-request memory allocation or a redundant lookup. No behavior changes.

What this removes from the hot path

• No heap allocation per header check: is_hop_by_hop_header used to allocate a lowercased String for every comparison; it now matches HeaderName constants directly, so the allocation is gone.
• ~10 fewer hashmap lookups per request: in-flight Gauge/Counter handles were resolved via get_or_create on every request; they're now resolved once and cached on ProxyHttp.
• No metric lookup on the user-agent path: the handle is cached per worker in an scc::HashMap<String, Counter>, so the warm path is a plain &str read instead of a metric-family lookup plus allocation.
• connection_info() called once instead of 5×: bound a single time per request and reused.
• No per-header revalidation when appending: headers_mut().append(...) skips reparsing each header name through the builder.
• No re-parse on response forwarding: dropped HeaderName::from_str. Backend header names are already valid, so the parse and its error path were dead weight.