Improve duplicate suppression hot paths and remove redundant outbound queue scans

[robekl]

This PR applies two focused performance improvements in packet processing and scheduling paths.

What changed

1) Packet-hash caching with faster duplicate-table prefiltering

• Packet hash computation is now cached on the packet object and reused when packet content is unchanged.
• The cache is explicitly invalidated at mutation/reuse boundaries so duplicate checks always reflect current packet state.
• Duplicate detection now uses small constant-size prefilters (for packet-hash prefix and ACK low-byte) to quickly reject obvious misses before any full table scan.
• Existing exact-match duplicate checks remain in place when a prefilter indicates a possible hit.

2) Outbound scheduler no-double-scan path

• The send loop no longer performs a separate “ready count” scan before selecting the next outbound packet.
• Scheduling now relies on a single retrieval pass, preserving existing timing, CAD, and transmit behavior.

Why this is beneficial

• Reduces repeated CPU work in high-frequency duplicate-suppression paths by:
  • avoiding repeat hash recomputation for unchanged packets
  • avoiding many unnecessary full duplicate-table scans
• Reduces scheduler overhead by removing redundant linear queue traversal.
• Keeps memory usage static and embedded-friendly with simple fixed-size structures.

Reviewer-focused implementation notes

• Duplicate prefilters are intentionally conservative:
  • They can produce false positives (which fall through to exact checks),
  • but do not introduce false negatives for duplicate detection.
• Hash cache correctness is maintained by explicit invalidation where packet fields are updated or packet instances are recycled.
• Behavior is unchanged for packet matching semantics and transmit gating; this PR targets execution cost, not routing policy.