Optimize Access Log and MQTT Session Paths

[LokiWager]

Summary

This PR combines three performance optimizations:

1. Access log formatter regex optimization (pkg/object/httpserver/mux.go)

• Move regex compilation to package-level variables.
• Avoid recompiling regexes during repeated formatter creation (for example, config reloads).

2. MQTT session loop optimization (pkg/object/mqttproxy/session.go)

• Reuse timestamp values from ticker.C instead of repeated time.Now() calls in the background loop.
• Reduce loop overhead from redundant system calls.

3. MQTT session pagination optimization (pkg/object/mqttproxy/broker.go)

• Add an in-memory session cache in Broker (map + sorted key slice).
• Keep cache synchronized through existing store watcher updates.
• Remove full-store fetch + full iteration from every paginated request.
• Make pagination deterministic via sorted keys.

Why

The previous implementation had repeated compute/system-call overhead and O(N) storage/CPU cost in hot paths:

• regex compilation during formatter construction
• repeated time.Now() calls in high-frequency loop
• full session scan + backend fetch on each pagination request

This PR reduces those costs while preserving behavior.

Measured Improvements (from original PR benchmarks)

• Access log formatter:

  • ~20473 ns/op -> ~12821 ns/op (~37% faster)
  • 112 allocs/op -> 74 allocs/op (~34% fewer)
  • 8758 B/op -> 4774 B/op (~45% less memory)

• MQTT session loop:

  • ~174.4 ns/op -> ~69.01 ns/op (~2.5x faster)

• MQTT session pagination:

  • ~206,380 ns/op -> ~3,917 ns/op (~52x faster)
  • eliminates O(N) backend getPrefix call per list request

Safety / Compatibility

• No intentional functional behavior changes.
• Broker cache access is protected by existing RWMutex usage.
• Existing pagination/filter behavior is preserved.