Configurable training

[saranggalada]

1. Added two new scenarios: BraTS and CIFAR-10
2. Pytrain now allows for highly flexible and secure training pipeline configurations
3. Fixed bugs in training code and ACI deployment
4. Restructured scenario subdirectories to demarcate different parties and components

[saranggalada]

Comments addressed and PR updated.

Summary of changes:

1. Introduced xgb_train.py workflow for XGBoost training and renamed existing train.py to dl_train.py to remove ambiguity.
2. New scenario "Credit Risk" that joins 6 datasets from 4 TDPs and trains using XGBoost with DP. Here, one of the TDPs (Bank A) also acts as the TDC, to reflect a more realistic scenario of collaboration.
3. Moved parsers/contructors for model, dataset, eval and loss configs into utilities
4. Minor change in policy.rego to allow single TDC container (if base model is being instantiated inside CCR)
5. Included 2 alternate model_config samples for COVID, one that trains using XGBoost, another that packages model as a SafeTensor file and not the default ONNX.
6. Updated ReadMes: i) Added mermaid diagrams to illustrate config creation for training. ii) Added info about datasets used in scenarios. iii) Modified main readme scenarios table. iv) Renamed ACR and Contract Service identifiers.
7. Added parameter "provider" in Join config to differentiate from dataset name, since single provider may have more than one dataset

To be done (in subsequent PRs):

1. "Create your own Scenario" guide and boilerplate automation
2. Classical ML training support
3. LLM Finetuning support
4. iSPIRT Contract Service 24*7 active demo, to provide users pre-signed contract sequence numbers for quick demos.
5. Optimize configurations / adopt existing standards (eg. Hydra / KubeFlow)