Fix fragile retry loop

[felipecrv]

1. RoundTripper contract violation (internal/client/client.go, errorHandler). On retry exhaustion it returned (resp, werr) with both non-nil. retryablehttp hands that straight to Go's net/http, which at net/http/client.go:262 logs "RoundTripper returned a response & error; ignoring response", discards the response, and leaks the connection — later surfacing as connection is already closed.
2. Fragile poll loop (connection.go, pollOperation). Long queries are driven by a 1s poll of GetOperationStatus. A single failed poll (idle timeout, reset, 5xx, EOF) made the sentinel abort the whole query — even though the operation handle is still valid server-side and the query keeps running. This is what "loses the statement handle."

Changes made:

• errorHandler now drains + closes the body and returns (nil, werr) — honoring the contract, matching retryablehttp's own default, and preserving the enriched error (reason-phrase headers).
• pollOperation now swallows transient GetOperationStatus errors and keeps polling, bounded by a new maxPollFailureWindow = 5 * time.Minute (so a genuinely-dead warehouse still terminates), and aborts immediately on context cancellation/deadline.
• Added tests: TestErrorHandler (nil response + body closed) and a TestConn_pollOperation subtest (recovers from a transient error instead of aborting).

Test verification

• TestErrorHandler, TestRetryPolicy, and all TestConn_pollOperation subtests pass (including the new resilience test).