fix(inference): Synchronize cpu_threads configuration for HPI predictors

[SankaraVenkatRam]

Summary

This PR fixes an issue where user-defined CPU thread settings (e.g. cpu_threads=2) were ignored when using the High Performance Inference (HPI) backend.

Although the value was correctly set in the legacy PaddlePredictorOption, it was never propagated to HPI, which instead used its own default (usually 10 threads). This could lead to unexpected CPU overuse, especially in containers or limited environments.

---

What changed

• Synced CPU thread settings between PaddlePredictorOption and HPIConfig via a new sync_threads step in BasePredictor
• Added support for the CPU_NUM_THREADS environment variable in:
  • OpenVINO and ONNX Runtime HPI configs
  • Default CPU thread handling in pp_option.py

---

Result

• Explicit cpu_threads values are now respected by the HPI backend
• CPU_NUM_THREADS provides a simple global override when no value is set in code
• More predictable and resource-friendly CPU usage

---

Why it matters

This aligns user intent with actual runtime behavior and avoids silent over-allocation of CPU resources.

[paddle-bot]

Thanks for your contribution!

[SankaraVenkatRam]

CI_XPU is failing during repository clone with a TLS handshake error on the XPU runner:
gnutls_handshake() failed

[SankaraVenkatRam]

The PR is still blocked due to the same CI failure on the XPU runner:
gnutls_handshake() failed during repository clone.

This appears to be a TLS/network issue on the XPU CI runner rather than something introduced by this PR.

[SankaraVenkatRam]

The CI failure is occurring during git fetch of the PR branch.

Logs show connection timeouts and proxy instability when accessing GitHub:

    Failed connect to github.com:443; Connection timed out
    
    RPC failed; result=52, HTTP code = 0
    
    The remote end hung up unexpectedly

This appears to be a CI runner network/proxy issue .

Kindly requesting maintainers to check the runner connectivity.

[Bobholamovic]

pp_option and hpi_config are primarily designed for two orthogonal features. The former is intended for configuring the Paddle Inference Engine (pp referring to PaddlePaddle), while the latter is used for High-Performance Inference (HPI) settings. Although pp_option can take effect when the HPI backend is set to Paddle Inference, I am concerned that updating the CPU thread configuration via pp_option in a way that impacts other backends (e.g., ONNX Runtime) may not be appropriate. Such behavior could potentially lead to confusion.

[Bobholamovic]

I would suggest adding the prefix PADDLEX_PDX to the environment variable.

[SankaraVenkatRam]

@Bobholamovic Thanks for the review! I've pushed a follow-up commit addressing both points:

1) Removed sync_threads() — each system now reads PADDLE_PDX_CPU_NUM_THREADS independently, keeping pp_option and hpi_config orthogonal.
2) Renamed env var to PADDLE_PDX_CPU_NUM_THREADS to follow the project convention

[Bobholamovic]

LGTM

[SankaraVenkatRam]

It looks like the CI_XPU job failed during the checkout step due to a TLS connection error when cloning the repository

[luotao1]

Sorry, there is something wrong with CI, could you please create a new PR to trigger CI?