fix(telegram): bound fallback-pool connection budget by pool count

[chelsealong]

What does this PR do?

Bounds the Telegram fallback-IP connection-pool budget so it no longer scales
multiplicatively with the number of fallback IPs.

Related Issue

Fixes #82678

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)

Root cause

TelegramAdapter.connect() builds two HTTPXRequest clients (general +
get_updates). On the fallback-IP path, each wraps a TelegramFallbackTransport
that owns one primary AsyncHTTPTransport plus one lazily-built transport per
fallback IP. The adapter passed the same httpx.Limits(max_connections=connection_pool_size)
(default 512, HERMES_TELEGRAM_HTTP_POOL_SIZE) straight into
TelegramFallbackTransport, which forwards that exact object to its primary
pool and to every fallback pool unchanged — that "caller limits win" behavior
is intentional and already covered by
test_caller_limits_win_over_pool_default / test_forwards_limits_to_inner_transports.

The result: the effective ceiling was 2 clients × (1 primary + N fallback) × 512, i.e. 1024 connections with zero fallback IPs and 2048 with one — far
beyond a typical 256 soft RLIMIT_NOFILE, before SQLite, logs, pipes, and
other sockets are counted. The fallback-IP list itself was also unbounded,
since DNS-over-HTTPS discovery returns every unique A record it sees.

#45507 lowers the same default from 512 to 64 but (as noted on the issue)
still doesn't bound the address count or the per-pool multiplication — with
one fallback IP it's still 2 × (1+1) × 64 = 256 before other descriptors are
counted, growing further with more addresses.

Changes Made

• plugins/platforms/telegram/telegram_network.py: added cap_fallback_ips()
  (bounds the fallback-IP list to DEFAULT_MAX_FALLBACK_IPS = 3, logging when
  truncated) and safe_fallback_pool_limits() (derates a httpx.Limits by the
  total pool count — num_ptb_clients * (1 + fallback_ip_count) — so the
  total connection budget across every pool stays close to what was
  configured, rather than multiplying by pool count).
• plugins/platforms/telegram/adapter.py: apply cap_fallback_ips() to the
  resolved fallback-IP list in connect(), and pass the derated limits (via
  safe_fallback_pool_limits()) into TelegramFallbackTransport instead of
  the raw per-client limits object. TelegramFallbackTransport itself is
  unchanged — its "caller limits win" contract stays intact.
• tests/gateway/test_telegram_network.py: added TestCapFallbackIps and
  TestSafeFallbackPoolLimits covering the cap and the derating math
  (including the zero-fallback-IP and near-zero-budget edge cases).

How to Test

1. bash scripts/run_tests.sh tests/gateway/test_telegram_network.py — new
   tests fail with AttributeError: module ... has no attribute 'cap_fallback_ips' / 'safe_fallback_pool_limits' against the pre-fix
   source (confirmed via git stash on just the two source files with the
   test file kept), and pass after the fix.

2. bash scripts/run_tests.sh tests/gateway/test_telegram_network.py tests/gateway/test_telegram_fallback_pool_release_71593.py tests/gateway/test_telegram_closewait_limits_31599.py tests/gateway/test_telegram_connect.py tests/gateway/test_telegram_network_reconnect.py — all pass:

   === Summary: 5 files, 45 tests passed, 0 failed (100% complete) in 6.6s (8 workers) ===
   

3. bash scripts/run_tests.sh tests/gateway/ (611 files):

   === Summary: 611 files, 5113 tests passed, 2 failed (100% complete) in 151.6s (8 workers) ===
   

   The 2 failures are both in tests/gateway/test_wecom_callback.py
   (test_build_event_extracts_text_message,
   test_poll_loop_dispatches_handle_message), unrelated to this change —
   they fail with the same AttributeError: 'NoneType' object has no attribute 'fromstring' in plugins/platforms/wecom/callback_adapter.py:400
   on HEAD~1 (before this commit) as well, confirmed by checking out the
   pre-fix source and rerunning that file in isolation.

4. ruff check plugins/platforms/telegram/adapter.py plugins/platforms/telegram/telegram_network.py tests/gateway/test_telegram_network.py — All checks passed!

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits
• I searched for existing PRs to make sure this isn't a duplicate (see
  Root cause section re: fix: stop Telegram gateway leaking sockets/FDs on reconnect storms #45507; fix(gateway): stop macOS gateway EMFILE — raise launchd fd limit + fix Telegram pool leak #39336 proposes raising the OS FD limit
  instead and is orthogonal)
• My PR contains only changes related to this fix
• I've run the test suite; all tests related to this change pass (2
  pre-existing, unrelated test_wecom_callback.py failures reproduce
  identically on HEAD~1, see step 3 above)
• I've added tests for my changes
• I've tested on my platform — not run against live Telegram infra; the
  new tests and existing pool/transport tests are unit-level with fake
  transports, no network

Documentation & Housekeeping

• N/A — no config keys added, no docs/architecture changes

Note on AI assistance

Drafted with AI assistance (an autonomous coding agent), with the diff,
tests, and RED→GREEN verification reviewed before pushing.