Drain the error queue and include the OpenSSL error details in
the exception for the default switch cases in handle_do_put_ssl_err,
do_handshake, do_get and do_shutdown.

These default cases are not expected to be reachable — the
SSL_get_error codes they cover require callbacks, modes, or BIO
configurations that seastar does not use. This change is defensive:
if they are ever reached, the error queue is now drained and the
OpenSSL error details are included in the exception message rather
than being silently discarded.

SSL_CTX_new can return a valid context while leaving errors on the
per-thread error queue. This happens when OpenSSL's system config
parsing partially fails but the failure is masked — for example,
ssl_do_config may call SSL_CONF_cmd which pushes errors (e.g.
SSL_R_NO_CIPHER_MATCH from an invalid Ciphersuites value in the
system openssl.cnf), but ssl_do_config itself returns success when
the system flag is set and conf_diagnostics is disabled.

Introduce a clear_stale_ssl_errors helper that drains and logs any
stale errors at debug level, and use it after SSL_CTX_new succeeds.

Several SSL operations can return success while leaving errors on
the per-thread error queue. This happens when OpenSSL internally
writes through our custom BIO (e.g. to send a close_notify alert,
a TLS 1.3 NewSessionTicket, or an application data record), our
bwrite callback fails with EPIPE and returns 0, but OpenSSL's
record layer misclassifies the failure as success.

Root cause: bio_write_intern (crypto/bio/bio_lib.c) passes our
bwrite return of 0 through unchanged. BIO_write returns 0. Then
tls_retry_write_records (ssl/record/methods/tls_common.c) checks
'if (i >= 0)' and classifies BIO_write returning 0 as success.
Our bwrite returns 0 following the documented BIO_write_ex contract
(1 success, 0 failure), but BIO_write's own contract says 0 means
"BIO is NULL or dlen <= 0" — not an error. The read side had this
same class of bug, fixed on master in OpenSSL commit be42447469.

The affected operations and how they reach our bwrite callback:

- SSL_shutdown: ssl3_shutdown -> ssl3_send_alert ->
  ssl3_dispatch_alert -> write_records -> tls_retry_write_records
  -> BIO_write -> bio_write_intern -> our bwrite callback
  (sending close_notify alert)

- SSL_do_handshake: state_machine -> TLS_ST_SW_SESSION_TICKET ->
  write_records -> tls_retry_write_records -> BIO_write ->
  bio_write_intern -> our bwrite callback (sending NewSessionTicket).
  Additionally, statem_srvr.c deliberately ignores flush failures
  via conn_is_closed() for this case.

- SSL_write_ex: ssl3_write_bytes -> tls_write_records ->
  tls_retry_write_records -> BIO_write -> bio_write_intern ->
  our bwrite callback (sending application data record)

- SSL_read_ex: can internally trigger writes (e.g. TLS 1.3 key
  update responses) through the same record layer write path.
  Not yet observed in test logs but covered defensively.

This is a workaround for the missing return-value translation in
bio_write_intern and can be removed once the upstream fix is
available.

OpenSSL's SSL_get_error relies on the per-thread error queue to
classify failures. In seastar's cooperative scheduling model, multiple
TLS sessions share the same thread. If one session leaves stale errors
on the queue and then yields, another session's SSL_get_error call may
misclassify the error — e.g. reporting SSL_ERROR_SYSCALL instead of
SSL_ERROR_SSL or vice versa — because SSL_get_error peeks at the
oldest error on the queue to decide the classification (ssl_lib.c,
ossl_ssl_get_error).

We have seen error queue contamination cause issues in practice
(e1625c8 "net: avoid propagating system errors to errno", cd02ecc
"ossl: Added ERR_clear_error if disconnected post write"). The current
code is believed to be correct — all error paths drain the queue
before scheduling points — but we add these checks to aid debugging
if a regression is introduced in the future.

Add verify_clean_error_queue() checks before every SSL_do_handshake,
SSL_write_ex, SSL_read_ex and SSL_shutdown call. If stale errors are
found, the first entry is logged at warn level and an assertion fires
to surface the problem in tests.