Replace TrinoResult generator with class-based iterator for error recovery

[srchilukoori]

Problem

Fixes #598.

TrinoResult.__iter__ uses a generator (yield). Python generators are permanently finalized after an unhandled exception propagates through the yield point — all subsequent next() calls raise StopIteration. Since #597 introduced lazy segment downloads during iteration, a transient I/O error (e.g. S3 timeout) mid-iteration kills the generator. fetchone() then returns None as if the query completed, silently dropping remaining rows.

Root cause

Generator frame finalization is a Python language constraint, not a bug in the iteration logic itself. Once an exception propagates through yield, the generator's frame is deallocated and cannot be resumed.

Fix

Replace the generator with a class-based iterator (__iter__ returning self, __next__ holding state in instance fields). The iteration logic is unchanged — batch prefetch via fetch() before exposing rows, same _rownumber tracking. The difference is that instance fields (_row_iter, _next_rows) survive exceptions, so callers can resume iteration after catching a transient error.

Testing

Added test_trino_result_iterator_survives_transient_error — injects a FailOnceIterator that raises IOError on the 3rd element, verifies the iterator resumes and returns remaining rows after the caller catches the exception.

Full unit test suite passes (123 tests, 0 failures).

[srchilukoori]

@hashhar , @azawlocki-sbdt I'm closing this PR as #615 is addressing all issues raised in the reviews above.