feat: add bounded TensorFlow CUDA E3 candidate

[rextio]

Summary

Adds a bounded TensorFlow CUDA E3 build-only candidate for the 0.1.2 integration branch.

The candidate lowers one frozen inference slice:

tf.matmul
→ tf.nn.bias_add
→ tf.nn.relu
→ tf.reduce_mean(axis=1)

Generated Rust calls a separate rextio_tensorflow_cuda_runtime over TensorFlow 2.21.0's TFE C API. It reuses already-resident cuda:0 float32 rank-1/rank-2 tensors and the existing eager context; it does not expose raw CUDA allocation, streams, executors, or host/device copies.

Frozen candidate contract

• Linux x86_64 GNU
• CPython 3.11
• Rust 1.93.1
• TensorFlow 2.21.0
• plugin API 1.6 / Device Provider API 1
• provider rextio-device-cuda
• capability cuda-tensorflow-tfe-linux-x86_64
• exact resource contracts:
  • framework.tensor
  • framework.eager-context
• inference/no-grad only
• inputs must already reside on canonical TensorFlow GPU:0
• fail closed when the authorized provider/capability, ABI symbols, owning DSO, device, dtype, rank, axis, or tape/forward-accumulator boundary differs

Evidence

• Adds a synthetic provider → Core generation → generated Rust cdylib link job on Ubuntu.
• GitHub Actions run 30067213294 passed all 14 component jobs plus ci-gate, including the CUDA E3 compile/link-only job.
• The hosted job does not import TensorFlow, load the extension, execute it, or execute CUDA.
• The generated inference body is checked for the exact four calls, strict order, and chained data flow.
• Machine-readable evidence is closed over the stable target, toolchain, provider, capability, operation, and resource-contract fields.
• The official TensorFlow 2.21.0 CPython 3.11 Linux wheel was inspected to bind DeviceList symbols to their actual owning image.

Explicit non-claims

• support_claim=false
• certification_ready=false
• real_gpu_evidence=false
• No performance claim
• No NVIDIA hardware execution, private-ABI runtime loading, numerical parity, or production support certification in this PR
• No Windows, macOS CUDA, non-x86_64 Linux, training/autograd, dynamic shapes, host tensors, implicit copies, multi-GPU, or fallback behavior

Real NVIDIA testing is intentionally deferred until suitable hardware is available. This PR does not target main, create a release, publish to PyPI, or change support status.

Local verification

• Full non-Cargo Python suite
• CUDA E3 focused tests
• Ruff
• mypy over src
• git diff --check
• package/sdist contract checks and generated Linux Rust cargo check were completed during candidate preparation