Fix stream reader readinto position accounting

[CAOShurong]

Fixes #295.

The C readinto() and readinto1() paths compress into a local output buffer,
but at EOF they increment bytesCompressed from self->output.pos. This can
leave tell() unchanged or wrap the counter after earlier output. Use the local
output.pos - oldPos instead, without changing compression or buffer ownership.

The regressions check cumulative emitted bytes after every call, bytes/stream
sources, small and multi-block inputs, mixed reads, guarded writable views,
repeated EOF and invalid/closed calls. They also verify decompression round trips.
The unchanged C backend fails two of the three new methods; CFFI passes them.

Validation on Windows x64, CPython 3.13.1, after building an sdist/wheel and
installing the wheel non-editably:

• pytest --import-mode=importlib <checkout>/tests/test_compressor_stream_reader.py -q:
  22 passed on C and CFFI, twice each.
• pytest --import-mode=importlib <checkout>/tests -q: C 251 passed/44 skipped;
  CFFI 227 passed/68 skipped, twice each. All 292 pre-existing test outcomes match
  the unchanged baseline.
• With ZSTD_SLOW_TESTS=1, compression-reader fuzzing: 16 passed per backend
  (--hypothesis-seed=295, default profile).
• ruff check ., changed-file format check, configured mypy (33 files), and
  git diff --check passed. C extension built with warnings as errors.

The setuptools test_suite/tests_require warnings are unchanged from baseline.
An empty bytes source stalled in the unchanged baseline and was excluded from
the additional API probe; an empty BytesIO source is covered by the regression.
Rust, other platforms/Python versions, and the full CI fuzzing profile were not
run locally.