refactor: review fixes and code improvements

- Add shared proxy module (bot/utils/proxy.py) to eliminate duplicate SOCKS5 parsing
- Fix AI client: escape HTML before Markdown→HTML conversion, unify timeouts,
  make health check optional and disabled by default, handle 429 retries
- Fix ai.py: correct forwarded message handling (aiogram 3.x forward_origin),
  pass relevant summaries via extra_system_content
- Fix dialogue.py: only respond to Astra's messages, use system context instead
  of prompt injection, answer on the limit message before phase transition
- Fix voice.py: load Whisper model in thread pool, safe WAV path generation
- Improve database.py: composite indexes, Boolean is_active, upsert file_id cache,
  add context cleanup helper
- Update weather.py and yadisk_download.py to use shared proxy connector
- Update yadisk.py: validate URL before cache clear, add download size limit,
  wrap sync file ops in to_thread
- Reuse S3 client via lru_cache
- Update setup_commands with /aiclear and /aiuser
- Update README, Dockerfile (Python 3.11), docker-compose (mount models)
- Pin dependency versions, remove unused httpx[socks]
- Add basic pytest tests for layout converter and voice normalization

bot/utils/s3_client.py

@@ -1,5 +1,6 @@
 import os
 import logging
+from functools import lru_cache
 
 import boto3
 from botocore.config import Config
@@ -10,6 +11,8 @@ logger = logging.getLogger("s3_client")
 
 ENDPOINT_URL = "https://storage.yandexcloud.net"
 
+
+@lru_cache(maxsize=1)
 def _build_s3_client():
     return boto3.client(
         "s3",
@@ -23,6 +26,7 @@ def _build_s3_client():
         ),
     )
 
+
 def upload_file(file_path: str) -> str | None:
     key_name = os.path.basename(file_path)
     try: