The refactoring in this commit is desirable to support shifting by adifferent `discard_size` (i.e. making that independent from`self.prev_start`).This commit also opportunistically adds an `assert!` that ensures thatshifting left won't accidentally clobber the immutable "lookback"window.

The refactoring in this commit is desirable because:* Main reason: To support calling the extracted function from another  place in a follow-up commit.* Secondary reason: To improve readability a little bit, by making  `fn unfilter_curr_row` slightly less noisy / more focused on its core  functionality, which is:    - extracting `prev_row`, and `row`    - calling `unfilter`    - updating `current_start` and `prev_start`

This commit makes the following two changes:* Falls back to unfiltering the current row out-of-place if an  `UnexpectedEof` error is encountered when in-place bytes cannot yet be  mutated.  Thisfixesimage-rs#639.  See the following changes:    - `test_partial_decode_success` in `tests/partial_decode.rs`    - Changes in `fn next_raw_interlaced_row` method of `Reader`      in `src/decoder/mod.rs`    - Changes in `src/decoder/unfiltering_buffer.rs`* Ensures that `finish_compressed_chunks` is always called (even if all  rows have already been successfully decoded).  This is required to  deterministically report errors, regardless of how the input is  buffered.  The other changes above would have resulted in fuzzing  failures without fixing this.  See the following changes:    - `test_partial_decode_with_unterminated_redundant_zlib_tail`      in `tests/partial_decode.rs`    - Changes in `fn finish_decoding` method of `Reader`      in `src/decoder/mod.rs`    - Changes of `src/decoder/read_decoder.rs`