perf(update): stream the target SHA-256 into the bspatch write

[peters]

Purpose

Autoresearch round 9 on the auto/update/ surface (continuing the
chain-apply optimization series #252 → #256 → #265). Phase
instrumentation in #256/#265 showed the per-step target SHA-256 is a
separate full read + hash pass after the bspatch write (~0.55 s
per step at the 1 GB benchmark scale). This change removes that
second read: the chunked bspatch write path now computes the
reconstructed file's hash while writing and returns it.

Behavior impact

• None observable. Same payload, same verification semantics:
  the streamed hash covers exactly the bytes written; mismatch errors
  keep the same message; the tamper-between-steps regression test
  still fails closed; per-step repack + full-archive SHA-256
  unchanged.
• API is additive: new chunked_bspatch_file_with_progress_and_sha256
  (+ ChunkedBspatchResult); the existing
  chunked_bspatch_file/chunked_bspatch_file_with_progress are
  untouched, so the archive strategy and all other callers are
  unaffected.
• Cost model: the eliminated phase was read (I/O) + hash (CPU). The
  hash CPU is now paid inside the write loop, so the net win is the
  saved read (~0.2-0.3 s/step at benchmark scale — larger on
  slow/flash storage where the re-read dominates).

Test evidence

• RUSTFLAGS="-D warnings" cargo test --workspace — all green
  (includes the new round-trip test asserting the returned hash
  equals the rebuilt file's SHA-256).
• cargo clippy --workspace --all-targets --all-features clean;
  strict panic-path clippy clean; cargo fmt --all -- --check clean;
  scripts/check-maintainability.sh pass.
• Same-session A/B (scale 1.0, 100 deltas, seed 42, 48-core shared
  host), interleaved base/new runs:
  • Phase-instrumented 10-step pairs (×4): the 540-563 ms/step
    target-hash phase is eliminated; per step ~2,270 →
    ~1,980-2,060 ms in 3 of 4 pairs (the 4th pair was a load-spike
    window that hit both sides' bspatch and repack).
  • 100-delta wall clock: new 186,122 / 167,182 / 157,514 ms vs
    base 178,113 / 166,808 ms. The ~0.2-0.3 s/step gain sits near
    this shared host's ~10% wall noise floor, so the claim rests on
    the phase-level evidence; documented in
    docs/performance/update-chains.md with the raw numbers.
• Download bytes unchanged (503,524 B); install tree identical
  (bench asserts).