xray-rust v0.4.1-rc.4

• Fixed an XHTTP/2 stream-up deadlock where a concurrent downlink poll could
  cancel the uplink's pending flow-control reservation and strand a sustained
  upload. Added deterministic split read/write coverage and an 8 MiB live
  Xray-core interoperability regression.
• Added safe pre-commit HTTP/2 GOAWAY retry handling plus regression coverage
  for non-replay after commitment and request/buffer ownership across
  cancellation.
• Accepted an HTTP/3 H3_NO_ERROR request reset only after every declared
  fixed-length request byte was delivered, while preserving failure for
  unknown-length or premature resets; added transport and full-load interop
  regressions.
• Added weekly broad pinned Xray-core interoperability and resource gates plus
  a warning-only upstream-main compatibility smoke check.
• Scoped the RC4 REALITY/Vision comparator omission to the measured stable
  sing-box v1.13.20 evidence: Xray-core v26.7.28's default minClientVer
  26.3.27 rejects that build's REALITY ClientVer 1.8.1. RC4 records
  --skip-sing-box and publishes xray-rust/Xray-core results only for those
  two workloads, while the generic harness retains sing-box REALITY support
  for compatible or patched binaries. The fixture and timeout remain
  unchanged.
• Published fresh five-run RC4 evidence against exact Xray-core v26.7.28 and
  stable sing-box v1.13.20: 139 validated series and 695 embedded results with
  exact revisions, binary hashes, a deterministic raw-archive digest, and new
  dated charts/tables. The publication records the reviewed sing-box
  gRPC/full-duplex/32 nondeterministic timeout and Xray-core H3 pressure/32
  reset/timeout boundaries without substituting isolated diagnostics.
• Made the benchmark publication depth-limit policy test portable across
  Python JSON decoders, and constrained JFrog-token secret-scan exemptions to
  the two reviewed dated benchmark replay records containing binary SHA-256s.
• Corrected roadmap, implementation-status, verification, configuration,
  migration, benchmark, and release documentation for RC4.
• Routing ip matchers are compiled once at load time into a shared sorted
  range index (xray-routing::IpRangeSet/IpMatcherSet), also used by DNS
  expectedIPs/unexpectedIPs filters. Outbound selection cost no longer
  grows with the size of a geoip: list (route-probe, 16 rules x 5000 CIDRs:
  ~152 us -> ~0.5 us per selection).
• Routing IP matching now follows Xray-core HeuristicIPMatcher exactly, as
  the DNS filters already did: IPv4-mapped IPv6 targets (::ffff:a.b.c.d) are
  matched as IPv4; an inverse matcher only applies to targets of an address
  family it has networks for, so !10.0.0.0/8 no longer matches an IPv6
  target; a doubly negated matcher is treated as positive (not expressible in
  configuration, so parsed configs are unaffected).
• CIDR prefixes longer than the address family width are rejected with an
  error at every construction site instead of being clamped.
• xray-bench route-probe --cidrs-per-rule N generates N distinct
  non-merging CIDRs per rule to expose per-matcher routing cost; 0 is
  rejected.