fix(warehouse-sources): a dropped source column auto-disables the schema instead of self-healing

Problem

A dropped or renamed source column auto-disables a SQL warehouse schema, and the message points at the wrong cause.

• All four SQL sources classify the driver's undefined-column error as non-retryable: Unknown column (mysql/source.py:302), does not exist (postgres/source.py:579, redshift/source.py:176), Invalid column name (mssql/source.py:111).
• external_data_job.py:493 calls update_should_sync(..., should_sync=False, ...) on one occurrence, so syncing stops until a person re-enables the schema.
• The MySQL message tells the customer to update the table's incremental field. That is not the cause when an ordinary data column was dropped.

Two paths reach it, and they differ in how long the failure lasts.

• An explicit column selection is never reconciled against the live catalog at sync time. prune_enabled_columns runs only in reconcile_source_schema_metadata (common/sql/base.py:203), which the API view calls on source reload. A saved column name that no longer exists is therefore named on every run, so the schema stays disabled until someone reloads the source.
• Sync-all resolves its projection from live discovery on the streaming connection (#95161). The same failure is possible only between that read and the query, and the next run would succeed.

The classification's own comment states the premise: the offending reference is persisted configuration, "almost always the configured incremental field", and "the streaming query reissues the same WHERE/ORDER BY on every attempt, so it fails identically forever" (mysql/source.py:292-301). That holds for the incremental field. It does not hold for a projection column.

Suggested fix

• Catch the undefined-column error before the first batch, rediscover the catalog once, rebuild the projection and the Arrow schema, then retry the read. The "no batch yielded yet" guard already exists per source: yielded_any in mysql.py, yielded in _stream_arrow_batches in redshift.py, and the offset == 0 guards in postgres.py.
• Prune an explicit selection against the rediscovered catalog on the same path, so a stale saved column stops being permanent.
• Keep the non-retryable classification for the configured incremental field, where the premise still holds.

Context

Raised by reviewhog on #95161. That PR shrinks the sync-all window from minutes to a couple of round trips. It does not address the explicit-selection path, and it does not change the auto-disable response.

Start with external_data_job.py:493 and the undefined-column handling in mysql/source.py, postgres/source.py, redshift/source.py, and mssql/source.py. Then trace explicit-column reconciliation in common/sql/base.py:203 and the pre-first-batch guards named in mysql.py, redshift.py, and postgres.py. Done means stale projection columns can self-heal without permanently disabling the schema, while configured incremental-field failures retain their non-retryable behavior.