PMK-2865: stop leaking Guzzle's exception hierarchy

[ewood-ac]

Jira: PMK-2865

Stacked on #168 (PMK-2061 / v8.0.0) — review that first. Ships in v8.0.0, not a later major.

Problem

processRestRequest() documented @throws \GuzzleHttp\Exception\GuzzleException and left the request call unguarded, making the HTTP client's exception hierarchy part of this SDK's public contract.

That coupling is total, not partial: because the SDK sets http_errors => false and maps responses to PostmarkException itself, the transport family is the only Guzzle family that can reach a caller. It's 100% of the leaked surface.

Guzzle 8 reclassified exactly that family, and #168 widens the constraint to allow it. Composer resolves highest, so upgrading moves callers onto Guzzle 8 whether they ask or not — and this silently stops matching:

try {
    $client->sendEmail(...);
} catch (\GuzzleHttp\Exception\ConnectException $e) {
    $retryQueue->push($message);   // no longer runs on a timeout
}

Change

Transport failures — DNS, connect, TLS, timeout, socket — become Postmark\Models\PostmarkTransportException, which extends PostmarkException, so existing catch (PostmarkException $e) blocks keep working unchanged. The original Guzzle exception is preserved on getPrevious().

catch (\Postmark\Models\PostmarkTransportException $e) {
    if ($e->isTimeout())           { /* at-least-once retry */ }
    if ($e->isConnectionFailure()) { /* safe to retry */ }
}

The two majors are less alike than they look

All version-specific knowledge lives in one private classifier. This was established by inspecting the installed packages, not by reading docs:

	Guzzle 7.15.2	Guzzle 8.0.2
ConnectException parent	TransferException	NetworkException
getHandlerContext() (the cURL errno)	present	absent
Timeout representation	ConnectException + errno 28	NetworkTimeoutException / ConnectTimeoutException / ResponseTimeoutException

An errno-based classifier would have silently degraded to "unknown" on Guzzle 8 — and that was caught because the tests run under both majors, not because it was predicted. The classifier now matches short class name first, falls back to errno, then to the message, and never references a class that exists in only one major.

Deliberately unchanged: http_errors stays false

Raised during review and rejected on two grounds. It would bypass the body parsing that produces Postmark's own ErrorCode/Message — every catch (PostmarkException) would get a generic Guzzle error with no API detail. And it would add Guzzle's response-exception family (ClientException/ServerException, also re-parented in 8) to the caller surface, which is the opposite of this change. http_errors => false is what has been insulating this SDK from Guzzle 8's loudest break.

Verification

Green under both majors — PHPStan [OK], 97 tests, 0 failures on 8.0.2 and 7.15.2. The Guzzle matrix added in #168 runs both.

New tests/TransportExceptionTest.php (credential-free) covers: wrapping, catchability as PostmarkException, classification without importing GuzzleHttp, the no-errno fallback, and that a 401 is still a plain PostmarkException — a response that arrives is not a transport failure.

Breaking change

catch (GuzzleException) around SDK calls stops matching. Documented in the v8 upgrade notes with the before/after. Anyone catching PostmarkException is unaffected.