iostream: Deprecate read_until_close and reimplement as a loop

[mokashang]

Fixes #3728.

BaseIOStream.read_until_close had its own dedicated code path through
BaseIOStream: a _read_until_close flag was set when the method was
called, special-cased in close() to deliver the buffered data (or let
_signal_closed fail with StreamClosedError(real_error=…)), and
coordinated with the internal _read_buffer as the data accumulated
there until the stream closed. The docstring already pointed callers
at the same loop a user can write themselves — read_bytes with
partial=True, iterated until a StreamClosedError with real_error=None.

This PR removes the special case and makes the method just that loop:

• read_until_close() kicks off the first read_bytes(..., partial=True)
  synchronously so an inline error (closed stream whose read buffer is
  empty, or a test that patches _try_inline_read) still raises from
  the call itself, matching the historical behavior
  (test_read_until_close_with_error).
• _read_until_close_loop() is a small async def that awaits the
  pending read, accumulates the chunk, requests the next partial read,
  and treats StreamClosedError with real_error=None or a
  connection reset as a normal close (returning the accumulated bytes)
  — preserving test_read_until_close_after_close,
  test_read_until_close_after_error_close, and
  test_read_until_close_after_connection_reset.
• Overflow is detected when len(buf) passes max_buffer_size and
  raised as StreamClosedError(real_error=StreamBufferFullError) with
  the same "Reached maximum read buffer size" log line used by
  _read_to_buffer. The old "error on read" log line from
  _handle_read is no longer emitted for this specific case because
  the overflow is now detected at the user-buffer level rather than
  inside the read handler; test_read_until_close_with_buffer_overflow
  is updated accordingly.
• The internal _read_until_close attribute and its branch in
  close() are deleted; the stale read_until_close mention in the
  _read_to_buffer_loop target-bytes comment is updated.

A new regression test (test_read_until_close_chunked) covers the
loop's reassembly path by driving max_buffer_size bytes through
read_until_close one byte at a time with intervening checkpoints,
ensuring the result fills up to (and no further than) the limit.

The method's docstring is now also tagged with .. deprecated:: 6.6
and a .. versionchanged:: 6.6 note explaining both the deprecation
reasoning (finer flow control, clearer error reporting after a partial
response) and the under-the-hood refactor.

Testing

• python3 -m tornado.test passes (1390 tests, 79 skipped).
• python3 -m unittest tornado.test.iostream_test.TestIOStream tornado.test.iostream_test.TestIOStreamSSL tornado.test.iostream_test.TestIOStreamSSLContext passes 108/108
  including every test_read_until_close_* case.
• black --check and mypy are clean on the changed files; flake8
  has no new warnings.

The changelog entry has been added under docs/releases/v6.6.0.rst
under both "Deprecation notices" and the new tornado.iostream
section.