Restore HistoryBuf write performance

[wyf-777]

Summary

• Fixes Performance regression for HistoryBuf(fer)::write() with heapless 0.9.1 #598.
• Restore HistoryBuf::write() optimization after the generic storage/view refactor introduced repeated backing-slice indexing that LLVM could not eliminate.
• Borrow the backing slice once, preserve the existing zero-capacity panic, and use the private write-position invariant to access the current slot without repeated bounds checks.
• Add a Divan benchmark based on the issue reproducer so the regression remains measurable.

Safety

write_at starts at zero and is reset before reaching capacity. The zero-capacity case is rejected before the unchecked access, and debug builds additionally assert write_at < capacity. The HistoryBuf test suite passes under Miri, including wraparound, drop-panic, and zero-capacity paths.

Performance

Windows x86_64, Rust 1.97.1, 8 MiB issue workload (cargo bench --bench history_buf):

• Before: 7.693 ms median, 1.09 GB/s median
• After: 223.1 us median, 37.58 GB/s median

Testing

• cargo test --features "alloc,mpmc_large,portable-atomic-critical-section,serde,ufmt,bytes,zeroize,embedded-io-v0.7" (294 unit, 8 concurrency, and 204 doc tests passed)
• cargo check --features defmt
• cargo test --release history_buf::tests (15 passed)
• MIRIFLAGS=-Zmiri-ignore-leaks cargo +nightly-2025-10-11 miri test history_buf::tests (15 passed)
• cargo +nightly-2025-10-11 fmt --all -- --check
• CI-equivalent Clippy for i686-unknown-linux-musl with all configured features
• Default and alloc checks for i686-unknown-linux-musl, riscv32imc-unknown-none-elf, armv7r-none-eabi, thumbv6m-none-eabi, thumbv7m-none-eabi, thumbv8m.base-none-eabi, and thumbv8m.main-none-eabi
• git diff --check

The Windows host cannot link executable tests with the embedded defmt feature because its runtime symbols are intentionally supplied by an embedded target; the feature itself passed cargo check, and the repository CI exercises the supported Linux configuration.

AI assistance

OpenAI Codex (GPT-5) assisted with issue analysis, implementation, benchmark construction, test execution, and pull request wording. The commit includes the corresponding Assisted-by trailer.