perf(decode): match prefetch 4-stage pipeline for long-distance matches

[polaz]

Context

execute_sequences reads from history at cursor - offset for each match.
For long-distance matches (offset > L2/L3 size), the source line is cold
in cache and the read stalls the pipeline.

Donor ZSTD_decompressSequencesLong (huf_decompress.c-adjacent path) uses
a 4-stage software pipeline: decode N+3 ahead of execute N, prefetching the
match source line 3-4 iterations in advance.

Goal

Port the 4-stage prefetch pipeline:

1. Decode sequence N+3 → emit prefetch for source line of match N+3
2. Decode sequence N+2 → continue prefetching
3. Decode sequence N+1 → source line for N now warm, execute match N
4. Steady state: every iteration prefetches one ahead, executes one behind

Acceptance criteria

• Pipeline triggered when match offset > prefetch threshold (~32 KiB or
  L2 line residency proxy)
• No regression on small-offset matches (short-offset path stays
  prefetch-free since data is already warm)
• Bench on at least one large-window corpus (decodecorpus-large-window
  or equivalent) shows measurable improvement
• All existing decode tests pass; no ratio regression

Files involved

• zstd/src/decoding/sequence_execution.rs (or wherever execute_sequences
  lives post-ci(deps): bump actions/github-script from 7 to 9 #133 interleave)
• zstd/src/decoding/decode_buffer.rs

References

• Donor: long-distance match prefetch in ZSTD_decompressSequencesLong

[coderabbitai]

🔗 Related PRs

#42 - perf(decoding): optimize sequence execution with overlap fast paths [merged]
#90 - perf(decoding): branchless offset history, prefetch pipeline, and BMI2 triple extract [merged]
#125 - perf(encoding): #124 Phase 4 — HC speculative tail check + MatcherStorage enum dispatch [merged]
#179 - perf(encoder): donor-parity greedy parse at L4 — ratio + speed win [merged]
#194 - perf(decode): fused sequence executor + SIMD/FSE hot-path cleanup + DoS-safe rollback [merged]

---

📝 Issue Planner

_{Check the box below or use the @coderabbitai plan command to generate an implementation plan and prompts that you can use with your favorite coding assistant.}

• Create Plan

---

🧪 Issue enrichment is currently in open beta.

You can configure auto-planning by selecting labels in the issue_enrichment configuration.

To disable automatic issue enrichment, add the following to your .coderabbit.yaml:

issue_enrichment:
  auto_enrich:
    enabled: false

💬 Have feedback or questions? Drop into our discord!