j2k 0.8.0

[0.8.0] - 2026-07-29

• Breaking: decoding is strict by default in j2k and j2k-native. Callers
  that intentionally accept the documented JP2/JPH optional-metadata
  recoveries must select DecodeSettings::lenient() per job; raw codestream,
  entropy, bounds, overflow, allocation, and resource-safety checks remain
  strict in both modes.
• Breaking: lenient decoding now warns only when a permitted recovery was
  actually used. J2kDecodeWarning::LenientDecodeMode is renamed to
  J2kDecodeWarning::LenientMetadataRecovery. New
  J2kView::parse_with_settings and J2kDecoder::new_with_settings
  constructors keep this policy local to one input.
• Adds CudaLosslessEncoder, a reusable, exclusively borrowed lossless-encode
  façade with explicit Auto, CPU-only, and strict-CUDA routing. Its opaque
  result reports requested and actual execution plus any unavailable-device or
  incomplete-route fallback; execution errors are returned without a hidden
  full CPU retry, and the encoder can be reused after failure.
• Routes 25–38-bit Auto lossless encode requests to the supported CPU path
  while preserving RequireDevice failure semantics.
• Defines stable, experimental, implementation, and binary API tiers in the
  ordered release manifest, and makes the 0.7.5→0.8.0 semver transition
  fail closed with an exact source- and behavior-break ledger. A later 0.8.x
  candidate must use a published v0.8.0 baseline.
• Fixes the published j2k-ml 0.7.5 cuda and metal clean-consumer
  failures. The adapters now use accelerator codec decode, explicit
  decoded-pixel readback, and ordinary Burn tensor upload through released
  dependency APIs.
• Adds CudaUploadBurnDecoder and MetalUploadBurnDecoder as explicit aliases
  for staged host-memory upload behavior while preserving CudaBurnDecoder,
  MetalBurnDecoder, and their inherent method paths unchanged.
• Removes the private CubeCL and wgpu patch sources from the release graph.
  This is not a direct-destination or zero-copy accelerator contract.