Add otel.sdk.metric_reader.collection.duration (Histogram),
otel.sdk.exporter.metric_data_point.inflight (UpDownCounter) and
otel.sdk.exporter.metric_data_point.exported (Counter) to ExportingReader.

ExportingReader accepts an optional MeterProvider to avoid a circular
dependency — pass a separate provider instance for self-observability.

- Add hasShutdown() guard in Logger::emit() before incrementing log counter
- Fix ExportingReader::countDataPoints() to use count() for arrays instead
  of iterator_count() which would exhaust generators
- Fix ExportingReader::doCollect() to record collection duration before
  export, not after, matching the spec intent
- Replace all literal attribute/metric name strings with OtelIncubatingAttributes
  and OtelIncubatingMetrics constants across TracerSharedState, LoggerSharedState,
  SpanBuilder, Span, BatchSpanProcessor, BatchLogRecordProcessor, ExportingReader

Replace iterator_count() fallback in ExportingReader::countDataPoints()
with 0 to avoid exhausting generators before export. Use '_OTHER' (semconv
catch-all) instead of 'export_failure' for metric export failures where no
exception class is available. Track dropped spans/logs in the processed counter
with error.type='_OTHER' to restore observability of queue-full drops. Add
tests verifying counter values after successful export and after queue-full
drops for BatchSpanProcessor, BatchLogRecordProcessor and ExportingReader.

- ExportingReader: wrap export() in try/finally so dataPointInflightCounter
  is always decremented even if the exporter throws
- ExportingReader: fix countDataPoints() to use is_countable() instead of
  is_array() (dataPoints is typed iterable, not array) and remove dead isset()
- ExportingReader/BatchSpanProcessor/BatchLogRecordProcessor: move
  ReflectionClass call inside the meterProvider non-null branch to avoid
  unnecessary reflection when self-observability is disabled
- ExportingReader/BatchSpanProcessor/BatchLogRecordProcessor: replace
  static \$instanceCount with spl_object_id(\$this) for stable, test-safe
  instance naming that does not accumulate global state across test runs
- BatchSpanProcessor/BatchLogRecordProcessor: remove dead \$dropped and
  \$processed fields; drops and processed counts are now fully captured
  by the self-observability counters
- Logger: add hasShutdown() check to isEnabled() to match emit() behaviour;
  previously isEnabled() returned true for post-shutdown loggers while
  emit() silently dropped records
- LoggerProviderFactory: forward \$meterProvider to LoggerProvider so that
  factory-created providers emit self-observability metrics
- Context: add withMeterProvider() to allow updating the config context
  after MeterProvider is constructed
- OpenTelemetrySdk: update \$context with the real MeterProvider before
  building span/log processors so that BatchSpanProcessor and
  BatchLogRecordProcessor receive it via \$context->meterProvider

Replace NoopCounter/NoopUpDownCounter/NoopHistogram defaults with
nullable properties and null-safe ?-> operator, eliminating all
deptrac violations where SDK depended on API\Metrics\Noop\*.
Also add Sum instanceof assertions in tests to satisfy PHPStan.

Replace early-return pattern with if/else so Phan correctly sees
each readonly metric property assigned exactly once per branch.
Also replace is_countable/count with is_array/count for dataPoints
to satisfy Phan's type narrowing, and remove unused Metric import.

- Change attribute array PHPDoc types from array<string, string> to
  array<non-empty-string, string> to satisfy psalm metric instrument
  iterable key constraints
- Add @psalm-suppress RedundantCondition for isset($metric->data) which
  guards against uninitialized typed property at runtime
- Fix test code: replace spread operator on iterable with is_array()
  assertion; replace reset()->value chain with explicit variable and
  assert($dp !== false) narrowing