fix: scope system-instruction change detection to a single agent

[mittalpk]

What does this PR do?

Fixes #462 — the "System Instruction Performance Analysis" warning in the dev UI false-positives on any multi-agent / sub-agent session, even when no system instruction was actually modified.

updateSystemInstructionFlags() in chat.component.ts flattens every LLM span across a session, sorts by time, and compares each chronologically-adjacent pair for a changed system instruction — with no partitioning by agent. In a root-agent-delegates-to-sub-agent flow, the LLM calls interleave across agents (e.g. root -> sub -> sub -> root), so two calls from different agents end up compared directly. Each agent has its own system instruction and its own cache context, so this is a category error, not a real change — independent of whether the two agents happen to share a model/provider.

The fix

Group the (already time-sorted) LLM spans by agent before doing the consecutive-pair comparison, instead of comparing across the whole session. The comparison logic itself is unchanged — it's the same per-pair check, just scoped to run within each agent's own sequence of calls.

Agent identity is resolved as eventData.get(span.attrEventId)?.author, falling back to the span's own attrAgentName (an OTel-promoted attribute that already exists on Span but wasn't used anywhere in the app) when the event lookup doesn't resolve — e.g. for spans without a matching event in eventData.

Testing

Added 3 unit tests for updateSystemInstructionFlags() (previously untested):

• Two LLM turns from different agents with different system instructions → no longer flagged.
• Two LLM turns from the same agent with a genuinely different system instruction → still correctly flagged (unchanged behavior).
• The attrAgentName fallback path, when the event has no resolvable author.

Confirmed the first and third tests fail against the pre-fix code (Expected true to be falsy) and pass after this change. Full test suite (646 tests) and ng build both pass.