Feat/tensor transport

[wyeth93]

feat(transport): add TensorMessage fast path

Summary

This PR adds TensorMessage, a transport-oriented message type for moving an
ordered set of CPU tensor buffers together with opaque metadata. It introduces
a direct TCP data path for compatible plain connections while preserving the
existing HTTP/2 packed-message path for TLS and explicitly configured legacy
compatibility modes.

The result is a lower-overhead remote tensor path without changing the existing
Message / RPC API for callers that do not opt in.

What changed

Core transport

• Add TensorMessage { version, metadata, buffers } and the associated
  encode/decode support.
• Add a raw TCP framing protocol for compatible plain connections (PTR1).
  A checked-out raw connection carries one request/response pair and is then
  returned to the pool; it is intentionally not multiplexed.
• Keep the HTTP/2 packed-tensor protocol (PTM1) for TLS and explicit HTTP/2
  compatibility modes.
• Allow the same listener to accept HTTP/2 and raw tensor connections.
• Use vectored writes and allocate the final receive buffers directly on the
  receive side; add pooling, size limits, and transport statistics.

Python API and dispatch

• Export pulsing.TensorMessage and pulsing.tensor_transport_stats() when
  the native binding supports them.
• Add actor-side receive_tensor dispatch and ActorRef.tell / ActorRef.ask
  support for tensor messages.
• Reserve receive_tensor from generic RPC dispatch so tensor payloads cannot
  accidentally take the normal message handler path.
• Keep imports safe for the RustPython binding: Tensor APIs are optional there,
  and attempting to convert a TensorMessage through that binding raises a
  clear unsupported error instead of failing an exhaustive Rust match.

Compatibility selection

Connection / setting	Tensor transport path
Plain connection with auto, raw, or unset transport	Direct TCP (PTR1)
TLS connection	HTTP/2 packed tensor (PTM1)
http2, legacy, or off	HTTP/2 packed tensor (PTM1)
Mixed-version deployment	Force http2 until all peers support raw transport

There is deliberately no silent fallback from a selected raw path to HTTP/2:
configuration and rollout policy should make the wire protocol unambiguous.

Semantics and limitations

• Tensor buffers must be CPU-resident and C-contiguous. Callers are responsible
  for staging CUDA tensors to CPU and making non-contiguous tensors contiguous.
• Pulsing treats metadata as opaque; it does not interpret dtype, shape, stride,
  or schema information.
• Local delivery can retain a read-only alias to the original buffer. Remote
  raw and HTTP/2 deliveries produce independent, writable receive allocations.
• The direct TCP path avoids assembling user buffers into one combined payload,
  but it is not network zero-copy: the data still traverses application-to-kernel
  and kernel-to-final-buffer copies.
• tell acknowledgement means transport enqueue, not handler completion or
  durability. Use ask with an application-level acknowledgement when that
  guarantee is required.
• This PR does not add GPU-direct transport, CUDA IPC, shared-memory transport,
  capability negotiation, automatic downgrade/retry, checksums, compression,
  encryption, request IDs, or streaming tensor responses.

Versioning and documentation

• Bump the participating workspace crates to 0.1.3.
• Update the Python and transport documentation/examples for TensorMessage
  and the transport-mode behavior.

Validation

Validated using /home/guoxu1/transferqueue_2/.venv and Maturin 1.11.5:

maturin develop --release --locked --offline \
  --pip-path /home/guoxu1/transferqueue_2/.venv/bin/pip

pytest tests/python/test_tensor_message.py -v
cargo fmt --all -- --check
cargo clippy --offline --locked --workspace \
  --exclude pulsing-py --exclude pulsing-bench-py --all-targets -- -D warnings
cargo test --offline --locked -p pulsing-rpymod
cargo test --offline --locked --workspace \
  --exclude pulsing-py --exclude pulsing-bench-py --lib --bins --tests --examples

The tensor-specific Python suite passes (12 passed), including raw transport,
HTTP/2 fallback, local lifetime behavior, actor dispatch, and reconnect cases.
The full Python suite also passes on a clean Codex-home configuration; the only
failure observed with a pre-populated local Codex cache is an unrelated existing
Forge plugin-catalog attribute mismatch from main.

[codecov-commenter]

Codecov Report

❌ Patch coverage is 58.65285% with 399 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
crates/pulsing-actor/src/transport/http2/server.rs	31.81%	120 Missing ⚠️
crates/pulsing-actor/src/transport/http2/client.rs	9.83%	110 Missing ⚠️
crates/pulsing-actor/src/actor/traits.rs	70.95%	79 Missing ⚠️
crates/pulsing-actor/src/transport/tensor.rs	82.73%	58 Missing ⚠️
crates/pulsing-actor/src/transport/http2/mod.rs	5.00%	19 Missing ⚠️
crates/pulsing-rpymod/src/bindings/message.rs	0.00%	7 Missing ⚠️
crates/pulsing-actor/src/system/handler.rs	0.00%	2 Missing ⚠️
crates/pulsing-actor/src/transport/http2/stream.rs	0.00%	1 Missing ⚠️
crates/pulsing-actor/src/watch.rs	0.00%	1 Missing ⚠️
crates/pulsing-rpymod/src/bindings/codec.rs	0.00%	1 Missing ⚠️
... and 1 more

Files with missing lines	Coverage Δ
crates/pulsing-actor/src/system_actor/mod.rs	69.54% <ø> (ø)
python/pulsing/__init__.py	78.02% <100.00%> (+1.01%)	⬆️
python/pulsing/core/__init__.py	76.08% <100.00%> (+2.27%)	⬆️
crates/pulsing-actor/src/transport/http2/stream.rs	80.93% <0.00%> (-0.28%)	⬇️
crates/pulsing-actor/src/watch.rs	92.89% <0.00%> (-0.48%)	⬇️
crates/pulsing-rpymod/src/bindings/codec.rs	0.00% <0.00%> (ø)
python/pulsing/core/remote.py	76.87% <92.85%> (+0.57%)	⬆️
crates/pulsing-actor/src/system/handler.rs	74.92% <0.00%> (-0.48%)	⬇️
crates/pulsing-rpymod/src/bindings/message.rs	0.00% <0.00%> (ø)
crates/pulsing-actor/src/transport/http2/mod.rs	84.58% <5.00%> (-4.10%)	⬇️
... and 4 more

... and 2 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.