feat: debounced partial spans — delay emission, cancel on fast finish

[PabloPardoGarcia]

Implements GLA2-244, stacked on #37 (merge that first; this PR's base is its branch).

What

init(partial_spans=True, partial_spans_delay=N) (env GLASSFLOW_PARTIAL_SPANS_DELAY, clamped [0, 60]s): pending snapshots are held N seconds and emitted only if the span is still open — a span finishing within the delay produces no pending on the wire at all. Default 0 is byte-for-byte the #37 behavior, with no scheduler thread even created.

This is the project's volume escape valve: most agent spans live milliseconds, so a few seconds of delay cuts pending traffic ~95% while anything worth watching live (open longer than the delay) still appears.

Design

• One shared scheduler thread regardless of span volume: deadline heap + (trace_id, span_id) → snapshot registry under a condition variable. schedule is O(log n), cancel is O(1) (heap entries for cancelled keys discarded lazily) — hot-path safe.
• The feat: emit partial (pending) spans at span start #37 seams did their job: emission wraps the existing _emit, cancellation rides the processor's already-received on_end, and the snapshot is still built at on_start and held — content set during the delay (set_input…) can never leak onto a pending. A delayed pending is byte-identical to an immediate one: zero wire/backend/UI changes.
• Fork-safe via the heartbeat sender's pattern (module-level register_at_fork over a weak set; children re-arm with an empty registry — parent spans aren't the child's).
• Lifecycle: shutdown() drops not-yet-due pendings (their final spans are flushing at that moment; anything emitted would be instantly superseded). One documented deviation from the ticket's prose (consistent with its ACs): force_flush() does not drop — flush happens mid-operation, and killing scheduled pendings there would silently disable liveness for spans that stay open.

Tests

TDD, 9 new: config default/env/clamp, scheduler pure logic with an injected clock (not-due, emit-once, cancel, shutdown-drops, post-shutdown no-op), one real-thread smoke test via a bounded Event.wait, and end-to-end through init() (fast span → zero pendings on the wire; delay=0 → #37 behavior). No sleeps anywhere. Full suite 165 passed; ruff + format + mypy --strict clean.

Per-kind delay (roots immediate, leaves delayed) is noted in the ticket as the future refinement; the scheduler supports it without wire changes.