fix(evaluation): isolate EvalCases per trace in tracing JSON import (#1021)

[klioen]

Summary

Fixes #1021: when BaseEvaluator.build_eval_set() imports a tracing JSON file containing multiple traces, the previous code accumulated all spans into a single conversation and a single EvalCase — mixing tool uses, session metadata (app_name/user_id) and user/model turns across independent sessions.

Changes

In veadk/evaluation/base_evaluator.py (_build_eval_set_from_tracing_json):

1. One EvalCase per trace — eval_cases, conversation, app_name/user_id and creation_timestamp are now reset for each trace_id group instead of being shared across the whole file.
2. Sort spans by start_time within a trace — the first/last call_llm spans now reliably map to that trace's user input and final output, even when spans are not ordered in the JSON file.
3. EvalSet timestamp — derived from the earliest generated case instead of whatever the last loop iteration left behind.

Test

Added test_tracing_file_multiple_traces_to_evalset covering:

• two traces → two isolated eval cases
• out-of-order spans within a trace (regression for the start_time sorting)
• session metadata, tool uses and conversations never crossing trace boundaries

3 passed in 1.76s

Fixes #1021