[AINode]: Integrate toto as a builtin forecasting model

[graceli02]

Summary

This PR integrates Toto (Time Series Optimized Transformer for Observability) by Datadog as a built-in forecasting model in AINode, following the same pattern as existing built-in HuggingFace models (sundial, chronos2, moirai2, timer_xl).

Toto is installed as an optional dependency via pip install toto-ts and loaded lazily at runtime, so it does not affect AINode's startup time or core dependencies when not in use.

Changes

• toto/configuration_toto.py — TotoConfig(PretrainedConfig) defining Toto's architecture parameters
• toto/modeling_toto.py — TotoForPrediction wrapper around toto-ts's Toto class, bridging ModelHubMixin-based loading with AINode's load_model_from_transformers mechanism
• toto/pipeline_toto.py — TotoPipeline(ForecastPipeline) implementing preprocess / forecast / postprocess with lazy imports guarded by a clear install message
• toto/__init__.py — package init with Apache 2.0 license header
• model_info.py — registered toto in BUILTIN_HF_TRANSFORMERS_MODEL_MAP with repo_id="Datadog/Toto-Open-Base-1.0" and auto_map pointing to the above classes
• AINodeTestUtils.java — added toto to BUILTIN_LTSM_MAP with expected state "active"

Design Notes

Toto uses huggingface_hub.ModelHubMixin rather than transformers.PreTrainedModel. TotoForPrediction.from_pretrained() delegates directly to Toto.from_pretrained() from the toto-ts package, then exposes a .backbone property (TotoBackbone) consumed by TotoForecaster in the pipeline. This avoids wrapping Toto in a standard HuggingFace PreTrainedModel while remaining fully compatible with AINode's existing loading infrastructure.

Setup Required

The Cluster IT - 1C1D1A / AINode test verifies that all built-in models report state "active" in SHOW MODELS. Since toto is newly added, its weights (~605 MB) must be pre-downloaded to the runner cache before the test will pass reliably.

Please run the following once on the self-hosted GPU runner before merging:

from huggingface_hub import hf_hub_download

hf_hub_download(repo_id="Datadog/Toto-Open-Base-1.0", filename="model.safetensors")
hf_hub_download(repo_id="Datadog/Toto-Open-Base-1.0", filename="config.json")

No code changes are needed.

---

This PR has:

• been self-reviewed.
  • concurrent read
  • concurrent write
  • concurrent read and write
• added documentation for new or modified features or behaviors.
• added Javadocs for most classes and all non-trivial methods.
• added or updated version, license, or notice information
• added comments explaining the "why" and the intent of the code wherever would not be obvious
  for an unfamiliar reader.
• added unit tests or modified existing tests to cover new code paths, ensuring the threshold
  for code coverage.
• added integration tests.
• been tested in a test IoTDB cluster.

---

Key changed/added classes (or packages if there are too many classes) in this PR

[CRZbulabula]

Hi Grace, this is your first PR (pull request) for Apache IoTDB repository, our community highly appreciate your contribution!

Next, let us talk about this PR. For timeseries foundation models' integration in AINode, we generally introduce their source codes then declare their open source license in the LICENSE file, which you can find in the project root dir.

Although installing the released package then invoke corresponding forecaster seems more convenient, this is usually not feasible in system engineering projects. To elaborate, different python projects often employ diverse versions of specified dependencies, resulting in package conflict. For instance, the transformers project updates its kvcache component significantly from v4.4x to v4.5x, meaning models build upon v4.4x cannot share the same dependency with models built on v4.5x.

To improve this PR, I suggest:

1. To trace the entire process of model forecast, dive into the forecast example scripts of Toto-1.0 model.
2. Integrate the Toto model through introducing the source code.
3. Packaging AINode, then verifying your integration via SHOW MODELS and SELECT * FROM FORECAST.

In addition to this process, you might encounter some problems in the model loading phase. This is because the Toto-1.0 model is loaded through ModelHubMixin interface, while the current version of AINode only accepts the PreTrainedModel format. Feel free to this challenge, we are integrating the PyTorchModelHubMixin interface and the corresponding PR can be refered soon.