ci: release 1.29.0

[kibae]

Summary

• Sync onnxruntime-server with upstream ONNX Runtime 1.29.0.
• Bumps version strings via deploy/update-version.sh 1.28.0 -> 1.29.0.

Upstream surface review (1.28.0 -> 1.29.0)

Header diff over include/onnxruntime/core/session/*.h plus the 1.29.0 release notes.

• No new element data types, enum members, or public headers. ONNXTensorElementDataType, GraphOptimizationLevel and ExecutionMode are unchanged, so value_info::type_name, parse_graph_opt_level and parse_execution_mode stay complete.

• No ONNX / opset bump. 1.28.0 moved to ONNX 1.22.0; 1.29.0 does not move it again, so sample-datatypes.py stays on opset 23.

• 4 new session-config key macros (kOrtSessionOptionEpEnableWeightless, kOrtSessionOptionEpContextSourceModelPath, kOrtSessionOptionsMaxShapeOverride, kOrtEpDevice_EpMetadataKey_WeightlessSupport) — nothing to implement: config_entries already forwards arbitrary keys to AddSessionConfigEntry and echoes the GetSessionConfigEntry readback.

• ORT_INTRA_OP_NUM_THREADS / ORT_INTER_OP_NUM_THREADS (#29688) are process env vars for default pool sizes. Not applicable — explicit intra_op_num_threads / inter_op_num_threads still take precedence, and we always set them explicitly when the caller asks.

• Version-pinned assumptions re-verified against 1.29.0 (commit a4d7ab7): packed uint2 initializers are still mis-loaded (the AllDataTypesModel tripwire did not fire, despite the packed sub-byte OrtApi::GetValue fix in #29157), and complex64/128 tensors still cannot be allocated. Both claims still hold; only the version named in the comments changed.

• POSIX telemetry is new in 1.29.0 (#27379) and is enabled in the official linux-x64 binary we ship against — documented in commit 47a2b33, images unchanged. Verified rather than assumed:

  • the 1DS collector URL https://mobile.events.data.microsoft.com/OneCollector/1.0 and its ingestion token are present as literals in the shipped libonnxruntime.so;
  • PosixTelemetry::enabled_ defaults to true (core/platform/posix/telemetry.cc:84);
  • it is suppressed only by ORT_DISABLE_TELEMETRY, ORT_RUNNING_UNIT_TESTS, or a CI env var (core/platform/telemetry_environment.h) — none of which a production container sets;
  • LogSessionCreation reports modelFileName, modelGraphName, producer, domain, weight type, graph/weight hashes, the full modelMetaData map and executionProviderIds.

  Since model and version are user-chosen path components under --model-dir, deployment-specific names leave the process by default. Decision: document, do not override. The published images keep upstream's default; the README options section and the Docker Hub page now describe the behaviour and the ORT_DISABLE_TELEMETRY opt-out. Baking ENV ORT_DISABLE_TELEMETRY=1 into the dockerfiles remains open as a separate policy change.

Test plan

• cmake build (Debug) and ctest pass locally against onnxruntime 1.29.0 GPU binary (9/9).
• Cross-platform CI passes (Linux / Windows / macOS).
• Docker images built and pushed: linux-cpu, linux-cuda12, linux-cuda13.
• Update Docker Hub repository description after merge: https://hub.docker.com/repository/docker/kibaes/onnxruntime-server/general