For a couple of releases we use collected IPs as a prioritized source
for connecting to Telegram. But apparently, they work way worse than it
should, and having connectivity to core ip ALWAYS gives better results.
Thus, this PR flips priorities, so users could have auto-update enabled
as a source of secondary addresses, not primary ones

- Replace manual make+copy with slices.Clone in Snapshot()
- Remove redundant _ = len(data); Snapshot() call alone is
  sufficient to exercise the lock under -race

- Move error check before Snapshot() to avoid unnecessary allocation
- Update existing tests to use Snapshot() instead of direct field access
- Add TestConcurrentAddSnapshot to explicitly exercise the mutex

1. Add sync.Mutex to ScoutConnCollected to eliminate data race between
   Add()/MarkWrite() in readLoop and learn() iterating results.
   Introduce Snapshot() for safe read access.

2. Increase bloom filter test size from 500 to 100000 to prevent
   false negatives from random eviction in the stable bloom filter.

3. Use Require().NoError() in TestHTTPSRequest to prevent nil-pointer
   panic on resp.Body.Close() when the request fails.

Fixes #425

This is a followup for https://github.com/9seconds/mtg/issues/412 it
makes sense to manage timers inplace instead of creating for new
goroutines: saves memory

- Use sync.Pool for relay buffers instead of stack-allocated arrays.
  A [16379]byte on the goroutine stack forces Go to grow it to 32KB
  (next power of two). Pooled buffers keep goroutine stacks small.

- Same fix for doppelganger write buffer ([16384]byte in conn.start).

- Replace idle goroutines with context.AfterFunc in proxy.ServeConn
  and relay.Relay. These goroutines existed only to wait on ctx.Done()
  and close connections. AfterFunc achieves the same without allocating
  a goroutine until the context is actually cancelled.

Net effect: at 3000 concurrent connections on a 1-vCPU/961MB VPS,
the unmodified binary drops 246 connections and falls to 10 MB/s.
With these changes: zero failures, 63 MB/s, 31% lower RSS.

Closes #412

Instead of a separate cert_probe.go that duplicates the scout's TLS
connection logic, measure the cert chain size directly from the same
HTTPS connections the scout already makes.

Changes:
- Extend ScoutConnResult with payloadLen field
- Add Write interception to ScoutConn for handshake boundary detection
- Scout.learn() now computes cert size (sum of ApplicationData between
  CCS and first client Write) alongside inter-record durations
- Ganger aggregates cert sizes across raids and exposes NoiseParams()
  via atomic pointer for lock-free reads from proxy goroutines
- Proxy reads NoiseParams from Ganger on each handshake instead of
  probing at startup
- Remove cert_probe.go, disk cache, and related config options
  (noise-cache-path, noise-cache-ttl, noise-probe-count)

Falls back to legacy 2500-4700 range until the first scout raid
completes (typically within 1-2 seconds of startup).

The hardcoded noise range (2500-4700 bytes) in the FakeTLS ServerHello
does not match the real certificate chain sizes of many popular fronting
domains (e.g., dl.google.com ≈ 6480 bytes, microsoft.com ≈ 13004 bytes).
This makes the proxy detectable by DPI systems that compare the
ApplicationData size with the real cert chain size for the SNI domain.

On startup, probe the fronting domain's actual TLS handshake size and
use the measured value ± jitter instead of the static range. Falls back
to the legacy 2500-4700 range if the probe fails.

Also adds optional caching of probe results between restarts
(noise-cache-path, noise-cache-ttl) and a configurable probe count
(noise-probe-count) under [defense.doppelganger].

Closes #408

Even if it makes sense to have a huge buffers, we do artificial delays
now. In that case we could achieve the same results with a lower buffer.
If not, then we won't send a packet bigger that this value