Stabilize PyTC v2 runtime and dual-schema config flows

[akgohain]

addressing two buckets of work:

1. Runtime/API correctness for training and inference.
2. Dual-schema config handling so legacy uppercase YAML and PyTC v2 lowercase YAML both behave in the client.

Runtime and API stabilization

• fixed training start so preset-loaded YAML is accepted, not just uploaded files
• fixed inference launch contract and added missing validation for config/checkpoint inputs
• added end-to-end inference status support
• added runtime log endpoints and in-app training/inference log panels
• preserved config origin paths so staged runtime YAML resolves relative _base_ includes correctly
• fixed proxy error handling so upstream worker failures propagate with accurate HTTP status and actionable detail
• added /training_logs and /inference_logs
• added /inference_status
• fixed /start_tensorboard proxying and missing logPath handling
• fixed tensorboard URL lookup to proxy the worker instead of returning a hardcoded value
• fixed checkpoint selection to use file mode

Dual-schema YAML support

• added schema-aware config helpers for legacy uppercase and PyTC v2 lowercase layouts
• updated training/inference path mutation logic to write to the active schema instead of forcing one layout
• updated architecture and slider mutations to respect the active schema
• disabled unsupported controls instead of silently writing ineffective fields
• synced current UI input/output selections back into the active YAML before launch so the posted config is not stale

Local startup UX

• made scripts/start.sh reuse healthy existing services instead of blindly starting duplicates
• redirected startup service output into .logs/start
• added cleaner readiness/error reporting during local startup
• made the data server fail gracefully when port 8000 is already in use