fix(tracing): flush spans per turn

[casey-brooks]

Summary

• flush OTel spans after each agent turn
• pass flush function through server for per-turn tracing reliability

Testing

• go vet ./...
• go test ./...

Issue

• #114

[casey-brooks]

Summary

• pass tracer flush hook into the server and force a flush after each turn
• update buildAgent plumbing and server tests for the new flush hook

Testing

• go vet ./...
• go test ./...

[casey-brooks]

Test & Lint Summary

• go vet ./... (no issues)
• go test ./... (passed: all, failed: 0, skipped: 0)

[casey-brooks]

Summary:

• enabled parallel tool calls for response requests to keep tool calls in a single LLM turn.

Tests:

• go vet ./...
• go test ./...

Lint:

• go vet ./...

[casey-brooks]

Summary:

• added tool-call batching instructions so the model emits all required tool calls in one response.

Tests:

• go vet ./...
• go test ./...

Lint:

• go vet ./...

[casey-brooks]

Summary:

• record llm.call spans only for the first model call and the final response to coalesce intermediate tool-only calls.

Tests:

• go vet ./...
• go test ./...

Lint:

• go vet ./...

[casey-brooks]

Summary:

• coalesced llm.call spans to count only the first tool call and final response.

Tests:

• go vet ./...
• go test ./...

Lint:

• go vet ./... (no errors)

Test results:

• passed: all
• failed: 0
• skipped: 0