perf(read): optimize reusable buffers and frame extraction

Separate leftover storage from reusable read chunks and add fast paths for contiguous frame extraction. Reclaim consumed leftover prefixes and avoid unnecessary read-buffer growth.

Keep framing in a unified loop, cover short-read buffer reuse, and clarify that BufReader fill_buf already refills through read_buf.

Author: Yunuuuu

benches/README.md

@@ -17,6 +17,11 @@ By default, `FramedRead` and direct `BytesMut` use `Read::read` with initialized
 destinations. `FramedRead` tracks initialized spare space; direct `BytesMut`
 initializes its tail on each refill. With `nightly`, both use `Read::read_buf`.
 
+The `BufReader::fill_buf()` comparisons use the same code in both configurations.
+On the tested standard library, `fill_buf()` already refills its internal buffer
+through the source's `Read::read_buf`; this does not depend on this crate's
+`nightly` feature.
+
 Read compares `FramedRead`, `BufReader`, and direct `BytesMut` buffering.
 `FramedRead` retains partial frames separately and refills whole chunks.
 `decode_view()` joins crossing frames; `decode_owned()` copies the selected

benches/read.rs

@@ -3,7 +3,8 @@
 //! 500 B, 1000 B, and 1500 B respectively, including LF.
 //! A fourth read-only workload uses Items larger than the initial buffer.
 //! Run with `cargo bench --bench read`; add `--features nightly` on nightly
-//! Rust to use Read::read_buf.
+//! Rust to use Read::read_buf for FramedRead and direct BytesMut buffering.
+//! BufReader::fill_buf already uses read_buf internally on the tested std.
 
 #![cfg_attr(feature = "nightly", feature(core_io_borrowed_buf, read_buf))]