Reduce SQL Server queue length monitoring query volume

[ramonsmits]

Spike addressing high SQL query volume from queue-length monitoring (one IF EXISTS + max-min(RowVersion) statement per tracked queue, every 200ms).

Changes

• Single query. Read approximate row counts for all tracked queues from sys.partitions (one query per catalog) instead of probing each queue table. No queue-table reads, no locks, only metadata visibility required. Accuracy is comparable to the existing max-min(RowVersion) estimate (both approximate; the catalog counter avoids identity-gap over-counting).
• Concurrent pacing. The poll interval now runs concurrently with the query, so the effective cadence is max(interval, queryDuration) instead of interval + queryDuration. That additive query-time term is what let the cadence drift past the 1s monitoring bucket and starve buckets of samples — the SQL Server and PostgreSQL can indicate report 0 as the queue length value #4556 false-zero "sawtooth". Removing the drift supersedes the SQL Server and PostgreSQL can indicate report 0 as the queue length value #4557 200ms oversampling workaround and lets the cadence return to a sane default.
• Configurable interval via QueueLengthQueryDelayInterval connection string part (default 1s, matching the finest monitoring bucket: 1-minute history / 60 = 1s per point), following the existing ASB convention.
• Adaptive back-off up to QueueLengthQueryMaxDelayInterval (default 10s) while every monitored queue is empty; now on by default. The base interval is always used while any queue has work, so the fix for SQL Server and PostgreSQL can indicate report 0 as the queue length value #4556 is preserved. Set the max equal to the base to disable back-off.
• SqlTable cleanup. Identifier properties renamed to Unquoted* to make the quoted-vs-unquoted contract explicit, converted to a primary constructor, and the duplicated per-table length-query branches collapsed into helpers.

Permissions

The bulk query reads only catalog views, so it needs less permission than the per-queue probe it replaces:

• Relies on standard metadata visibility — any permission the monitoring user holds on a queue table (e.g. SELECT) makes that table's row visible. No VIEW DATABASE STATE required (hence sys.partitions rather than the sys.dm_db_partition_stats DMV).
• The previous query read table data (SELECT max([RowVersion]) … FROM <queue>), which already requires SELECT on each queue table — strictly more — so any user that worked before continues to work.
• A queue table the user has no permission on simply returns no row and is treated as "does not exist" (same effective outcome as the old IF EXISTS → -1).

[mauroservienti]

One minor comment, @ramonsmits. Otheriwise, it's good to go.

[johnsimons]

LGTM