fix(evaluator): stop concurrent Evaluator workers sharing one Agent's short_term_memory

[AmirF194]

Root cause

Evaluator._create_new_agent_manager() (per-thread, used by _parallel_evaluate)
and Evaluator._async_execute_workflow_graph() (per-task, used by
async_evaluate) each build what their own comments call a "new"
AgentManager from AgentManager(agents=self.agent_manager.agents, ...).
That list holds the same Agent instances as the original manager, so every
concurrent worker runs the real Agent.short_term_memory (a 5-message
sliding-window deque, written by _prepare_execution/_create_output_message
on every call) against one shared object with no lock. Two benchmark
examples evaluated at the same time end up reading each other's messages
back as their own LLM context.

Fix

Added _agents_with_fresh_short_term_memory(), which clones each agent with
model_copy(update={"short_term_memory": ShortTermMemory()}) so every
worker gets its own memory while the clone still shares the agent's llm,
actions, and tools with the original (the docstring on
_create_new_agent_manager already said "same configuration but new locks";
this makes the per-agent mutable state match that intent too). Used at both
call sites.

Verification

• New test reproduces the bug through the real, unmodified
  Evaluator._create_new_agent_manager(): two threads each write 5 messages
  into their own "independent" manager's agent and read back a contaminated,
  interleaved context. Fails on main (the two managers hand back the same
  Agent object); passes on this branch.
• A second unit test exercises _agents_with_fresh_short_term_memory()
  directly (the same helper the async call site uses), asserting each clone
  gets its own ShortTermMemory while llm/name stay shared.
• tests/src/evaluator/ (9/9) and tests/src/agents/ (69/69) pass on the
  branch; ran both suites against an unmodified checkout of this repo's
  requirements.txt dependency set plus a stub for the heavy optional tools
  (browser/vector-store backends) this code path never touches.
• Did not run the async path (_async_execute_workflow_graph) end to end
  through a full WorkFlow; it calls the same now-tested helper, but that
  call site itself is covered only by the unit test above, not a live
  concurrent-workflow run.

Fixes #273