Metrics: add histogram bucket views per metric-name pattern

[1fanwang]

OTel histogram metrics in Airflow share one bucket layout regardless of what they measure, so a metric whose useful range differs from the default gets uninformative tails — task duration buckets fine, but scheduler-loop duration needs finer low-end resolution, and second-to-hour delays need coarser high-end. #64207 set an exponential default per instrument type, but non-timer histogram families (*_count, *_duration, *_delay) still resolve their boundaries at each call site, so the same family can end up shaped differently depending on which module created the instrument.

This adds a declarative metric-name pattern → bucket aggregation map in shared/observability and layers the resulting OTel Views on top of the existing instrument-type baseline. Latency families get exponential buckets, counts get a small linear range, delays get a wide range; the per-instrument-type default still applies and the pattern views only refine it for matching names. Deployments needing a different layout pass an override dict.

Closes #66801

Tests

New test_histogram_buckets.py covers the default pattern map, per-pattern aggregation resolution, the custom-mapping override, and that one View is built per entry. test_otel_logger.py is updated to assert the layered shape (baseline view followed by the pattern views).

before/after on the discriminating test

Reverting otel_logger.py to upstream/main (baseline view only, no pattern views):

FAILED ...test_get_otel_logger_uses_exponential_histogram_view
    assert pattern_names == {"*_count", "*_duration", "*_delay"}
E   AssertionError: set() != {'*_count', '*_delay', '*_duration'}

With the change restored: 10 passed (full test_histogram_buckets.py plus the updated otel_logger assertion).

A standalone MeterProvider(views=build_views_for_patterns()) driving an InMemoryMetricReader confirms the SDK resolves per name end to end: task_duration collects as ExponentialHistogramDataPoint, while schedule_delay / retry_count collect as fixed-boundary HistogramDataPoint. Without the patch all three fall through to the exponential default regardless of name.