Add CUDA device-pointer plane_codec ABI with explicit selection

[khwstolle]

Summary

Adds a device-pointer plane_codec ABI (encode and decode) so a
GPU-resident codec can operate on device memory directly, supplied
entirely by a third-party plugin. The new symbol pair mirrors the
existing hardware_backend_v1_* CUDA ABI but attaches to the
plane_codec kind, so it participates in container framing the same
way a built-in codec does.

Why

The existing plane_codec dispatch path assumes a contribution's
data already lives in host memory. A GPU-resident codec needs its
compressed and decompressed buffers to stay on-device: routing every
plane through a host round trip would erase most of the reason to use
a GPU codec at all. This adds the missing piece, a symbol pair that
takes device pointers directly, plus the plumbing to reach it from
Rust, Python, and the CLI.

What changed

ABI and dispatch (Rust core)

• Two new CUDA C-ABI symbols (encode, decode) attached to the
  plane_codec kind. Encode reports its required output size
  explicitly, since the existing decode_rc error code cannot
  round-trip an arbitrary byte count.
• A DispatchedPlaneCodecCuda trait and native adapter bridge the
  ABI, resolved through a capability-gated router mirroring the
  existing hardware-backend router's caching and hardware_class
  gating.
• Third-party encode becomes reachable when a codec is explicitly
  selected by name or id. It stays unreachable through the automatic
  trial-encode search, whose return shape the flat third-party ABI
  cannot express.
• A name/id resolver (resolve_codec_selector) turns a human-typed
  selector into a canonical id, erroring loudly and naming every
  candidate on an ambiguous or unknown name, rather than picking one
  silently.
• A Rust-only integration test pins that decoding through this router
  never needs the Python bindings.

Python bridge and surface

• A PyO3 bridge exposes the encode/decode pair to Python, sharing
  DLPack tensor parsing, a buffer-protocol fallback, and
  device-ordinal agreement checking with the existing hardware-backend
  bridge rather than duplicating them.
• CompressionConfig/DecompressionConfig gain codec/device
  fields; compression gains a guard against a mismatch between an
  explicit device and an input tensor's own device.
• The CLI gets --codec/--device flags, wired through every
  config-construction path each command reaches.
• Setting codec or device on a config with no dispatch path to
  consume it raises immediately, naming the field, instead of
  silently falling back to the automatic codec search or CPU-only
  decode.

Testing

• cargo test --workspace: all passing, clippy clean.
• pytest across the full suite (maturin develop build): all
  passing; CUDA-dependent paths skip without a GPU.