[Tools] Add PID feature extractor and ONNX inference tasks

[forynski]

Adds two new tasks under Tools/PIDFeatureExtractor/:

• pidFeatureExtractor.cxx: extracts a 34-feature / 7-detector-group PID
  feature set (TPC, TOF, TRD, ITS, EMCal, HMPID, centrality) from AO2D
  data for both MC (reconstructed + truth) and real/raw data, writing a
  ROOT TTree (and optionally CSV). Includes optional DPG-style track
  cuts (off by default) and an optional Bayesian PID posterior for
  comparison against the ML model.

• pidOnnxInference.cxx: runs a trained ONNX model over the extractor's
  output and writes per-track class probabilities, using
  o2::analysis::MlResponse (Tools/ML) for CCDB/local model loading.

Notes for reviewers:

• pidFeatureExtractor.cxx writes a plain TFile/TTree rather than an AOD
  table (DECLARE_SOA_TABLE/Produces<>). This is deliberate: two
  Produces<> tables sharing a column prefix in one struct triggered a
  reproducible compile failure against the current O2 build
  (ASoA.h/MetadataTrait constraint errors). Manual ROOT I/O avoids it.
• pidOnnxInference.cxx is a plain DataProcessorSpec, not an
  AnalysisTask - it has no AOD subscription (it reads the extractor's
  output file directly), and runs standalone after the extractor
  rather than chained in the same pipeline.
• The model's expected input feature order (in pidOnnxInference.cxx) is
  a reasonable default but not yet verified against the actual training
  code - flagged in the README.
• modelPathsCcdb currently holds a placeholder CCDB path; needs a real
  location once the model is deployed.

See README.md in the new folder for full usage and configuration
details.

[github-actions]

O2 linter results: ❌ 0 errors, ⚠️ 14 warnings, 🔕 0 disabled