Refactor common result metadata processing into a shared helper

[cquil11]

Problem

utils/process_result.py and utils/agentic/aggregation/process_agentic_result.py independently parse and validate common benchmark metadata. PR #2100 adds heterogeneous prefill/decode hardware metadata to both paths, which highlights the maintenance cost and risk of the two implementations drifting.

Goal

Move shared environment parsing, topology validation, and common result fields behind one reusable implementation consumed by both regular and agentic result processors.

Acceptance criteria

• Regular and agentic result processors use the same code for common metadata and multinode topology fields.
• Optional PREFILL_HARDWARE / DECODE_HARDWARE handling is implemented once, with both-or-neither validation.
• Existing homogeneous, heterogeneous, single-node, and multinode result schemas remain compatible.
• Shared table-driven tests cover the common behavior; processor-specific tests focus on integration and metric output.

[cquil11]

Suggested implementation:

1. Add a pure helper module such as utils/result_metadata.py. Have it accept a Mapping[str, str] rather than reading os.environ directly.
2. Introduce build_common_result_fields(env, scenario_type=...) for shared fields such as runner hardware, model, framework, precision, speculative decoding, disaggregation, and scenario type.
3. Introduce build_topology_fields(env) returning both serialized topology fields and derived values (num_gpus, effective TP/EP, and DP attention). This function should own single-node versus multinode parsing, prefill/decode worker fields, and paired optional hardware validation.
4. Keep metric-specific work in the current processors: process_result.py handles serving latency/throughput conversion, while process_agentic_result.py handles request/server metric nesting.
5. Move the topology cases into table-driven tests against the shared helper: single-node, homogeneous multinode, heterogeneous multinode, prefill-only hardware, and decode-only hardware. Leave one integration smoke test in each processor to verify that shared metadata reaches its final JSON.

A small immutable dataclass (for example, TopologyMetadata) would make the derived values explicit and prevent the processors from recomputing counts differently. The dataclass can expose as_result_fields() so output naming also lives in one place.