feat(otel): Give each session's cycle its own reactive_update span

[schloerke]

Stacked on #2520 (on #2517 → #2515 → #2508). Fixes #2508's TODO: reactive_update spans losing session.id.

This branch also contains a merge of the current #2517 into #2520's branch, which #2520 doesn't have yet. Until #2520 merges #2517, this PR's diff shows those changes too; the change itself is the last commit.

Why session.id was lost

On main, every flush ran inside a session's message loop, so reactive_update picked up the session that triggered it, even though the flush is global and runs every session's effects. With #2508, a requested flush runs in a fresh contextvars.Context() on purpose, so it carries no session at all. It can't carry the requesting one either: requests from several sessions are merged into one flush. And with concurrent effects, the flush span ended as soon as effects had started, before their async parts finished.

Change: follow Shiny for R

R ties the span to the session's cycle, not to the flush (R/otel-reactive-update.R; started in incrementBusyCount(), ended in decrementBusyCount()). #2508 gave py-shiny the same per-session busy count, so:

• Start: AppSession._increment_busy_count() going 0 → 1 starts the session's reactive_update span, with session.id. Its parent is the current span, so the first cycle sits under session_start; later cycles each start a trace of their own.
• End: _decrement_busy_count() going back to 0, or the session ending, ends it. It covers the session's async effects to the end.
• Children: Effect_._run (and so every render output) runs under its session's span, so effect, output, and calc spans are its children.
• The flush has no span of its own anymore. Effects without a session have no reactive_update parent.
• New helpers: start_otel_span() / use_otel_span() in shiny/otel/_span_wrappers.py for a span that isn't one block of code.

Docs

• shiny/otel/__init__.py collection-level docs.
• The bundled skill's otel reference: the hierarchy diagram and level table.
• A CHANGELOG entry under Breaking changes.

Verification

• New tests/pytest/test_otel_reactive_update.py (9 tests):
  • effect and calc spans are children of the session's span, which has session.id and sits under session_start;
  • two sessions in one flush each get their own span;
  • the span lasts until an async effect finishes;
  • one span per cycle, with later ones as roots;
  • session-less effects are roots;
  • a session ending mid-cycle ends its span;
  • the collection levels session / reactive_update / all.
• Tests for the old flush-level span are replaced by one asserting that a flush creates no reactive_update span.
• Removing the effect's use of the span, or the end at idle, fails the new tests.
• pytest tests/pytest: 1332 passed. Six parallel runs, at one worker per core, all passed. pyright, pyrefly, flake8, black, and the skill validator are clean.