feature: cross-run pattern detection and anomaly alerting

[georgemarselis-nvi]

Once sufficient run history exists in the PostgreSQL database (#162), implement
a pattern detection layer that watches for anomalies across runs rather than
within a single run.

Examples of patterns worth detecting:

• a specific stage failing repeatedly across runs
• a specific sample consistently missing R2
• file size outliers occurring at a higher rate than baseline
• NIRD quota consumption rate trending toward exhaustion
• delivery times increasing over time (infrastructure degradation signal)
• runs from a specific instrument serial number producing worse QC metrics

Implementation notes:

• requires sufficient run history to establish a baseline - not useful until
  at least 3-6 months of data exists in feature database: create a postgresql instance that will contain 2 databases: demux and shinylims #162
• anomaly thresholds should be configurable, not hardcoded
• alerts via Teams (feature request: integrate demultiplexing script with Teams chatbot. #145) and email (feature request: Email notifications #24)
• this is a read-only layer on top of the operational database - no writes,
  no side effects

Depends on: #162 (PostgreSQL), #145 (Teams notifications), #200 (file size
statistics as input signal)

Related: #139 (detect when a single run has not completed in due time).
#139 is per-run timeout detection; this ticket is cross-run pattern analysis.
Both feed into the same anomaly alerting layer.

Not in scope: machine learning, statistical modelling. Simple threshold and
trend detection only for v1.