[Feature] OpenTelemetry GenAI exporter for Agent Traces (Phase 1: Java, spans-only)

Start by reading the serialized JSON Event Log contract in #924 and the existing trace_tree.py diagnostic shape. Define the standalone Java converter around the issue's span, ID, grouping, and incomplete-execution rules, then verify that repeated exports are deterministic and malformed records produce the specified diagnostics.

Search before asking

• I searched in the [issues](https://github.com/apache/flink-agents/issues) and found nothing similar.

Description

Tracking issue for Phase 1 of the OTel exporter discussed in #929, as a follow-up to #900 and the Execution Event recording work in #924.

Scope (Phase 1):

• A batch, offline converter over already-collected Event Log files (event-log.trace.enabled: true). It maps Execution Events to OpenTelemetry GenAI spans: one inputRunId becomes one trace with a synthesized invoke_agent root, and the recorded execution hierarchy defines the rest of the span tree. Records are grouped by inputRunId, sorted by timestamp, and end-of-input marks completion.
• Span mapping follows #929: llm uses CLIENT; tool, action, parser, and the run root use INTERNAL. parser maps to parse {name} with gen_ai.operation.name: "parse".
• Incomplete executions are still exported rather than dropped. A start without a terminal becomes a zero-duration span with status UNSET and flink_agents.execution.incomplete = true, plus an INCOMPLETE_EXECUTION diagnostic. The inverse case uses MISSING_START.
• Diagnostics follow the existing trace_tree.py shape (code, id, message, file location), including MALFORMED_RECORD for invalid input.
• The input contract is the serialized JSON Event Log record, not a particular file name. Explicit files are read as-is, unknown fields are ignored, and directory inputs may discover events-*.log files for convenience. For multi-subtask jobs, completeness of the supplied file set remains the caller's responsibility.
• Framework IDs are deterministically mapped to fixed-width OTel trace/span IDs so repeated exports produce the same IDs. Delivery remains at-least-once.
• Configuration is provided through converter arguments such as endpoint, protocol, and service.name. Content capture is excluded from Phase 1.
• Delivered as an optional standalone module, not bundled into dist, with OTel dependencies aligned through the opentelemetry-bom.
• Because the serialized Event Log format is shared across runtimes, the same converter works for both Java and Python agents.

Out of scope, tracked separately:

• A durable/replayable Event Log sink and a continuous OTel consumer with checkpoint/replay semantics.
• Content capture and richer usage metadata, pending the entityMetadata work discussed in #929 / #924.
• Optional OpenInference classification (openinference.span.kind) for backends such as Phoenix/Langfuse.

Rejected alternative (Flink's native OpenTelemetryTraceReporterFactory): see the note in #929.

Depends on #924 (recording contract).

Are you willing to submit a PR?

• I'm willing to submit a PR!