Surface foreign ExecutionPlan metrics via FFI (FFI_MetricsSet)

## Background

`PyMetricsSet` currently wraps the native DataFusion `MetricsSet` only. When a foreign `ExecutionPlan` is imported through `FFI_ExecutionPlan` (for example a plan implemented in a separate Rust extension, Go, or another Python module), there is no FFI round-trip for metrics: ``plan.metrics()`` on such a foreign plan returns whatever the underlying implementation happens to populate, with no standard way for the foreign side to surface DataFusion-style metrics back across the boundary.

## Why deferred

There is no `FFI_MetricsSet` struct in the upstream `datafusion-ffi` crate at v54 to wrap; this is a pipeline-symmetry concern rather than a feature gap with a clear upstream target. No open user requests at the time of audit. Filed for tracking so that when upstream lands FFI metrics support, the Python bindings can pick it up immediately. The accepted path is probably: upstream adds `FFI_MetricsSet` and threads it through `FFI_ExecutionPlan::metrics`; the Python bindings then expose a `from_pycapsule` helper on `MetricsSet` analogous to the other `*Exportable` patterns.

Start by checking the upstream datafusion-ffi crate beyond v54 for an FFI_MetricsSet and its connection to FFI_ExecutionPlan::metrics. Then inspect the Python MetricsSet and existing *Exportable from_pycapsule patterns. The work is complete when foreign execution-plan metrics can cross the FFI boundary and be exposed by the Python bindings.

Escrita pelo modelo de indexação a partir do texto da issue.