SCRUM-1069-explore profile: detect cumulative and snapshot measures

[emnarahmouni-exmergo]

Closes : #219
explore profile --check-cumulative (#219): flags a numeric column that
looks like a running total, an account balance, or a point-in-time
snapshot instead of a per-row increment. This shape profiles identically
to an ordinary measure (same type, null fraction, uniqueness), so summing
it downstream is a common and severely damaging misreading that nothing
else in explore profile catches.

How it works

• Signal: within an entity (a repeating, id-shaped column that is not
  itself the table's key) ordered by a temporal column, a cumulative or
  snapshot measure almost never decreases from one observation to the
  next. A genuine per-row increment has natural ups and downs.
• The probe is a single aggregate-only, window-function query authored
  once in DuckDB SQL and transpiled per connector dialect via sqlglot
  (mirrors relationships.py's overlap-probe pattern). No row's measure
  value ever leaves the engine, only comparison counts.
• Priced and gated like --verify: the base profile always completes
  and saves first; the check runs as a second, individually skippable
  phase, dry-run priced only once profiling knows what to test, and can
  come back as pending_confirmation on a billed connector rather than
  discarding already-paid-for work.
• A table missing an entity key or a temporal column is skipped with an
  explicit note, never silently reported as clean.

Bugs found and fixed during manual testing

Two real bugs surfaced while exercising this against a live DuckDB file
(not caught by the initial test pass, since it always exercised a fresh
scan):

• A dataset served from the freshness cache is a deep copy, not the
  object the cache write carries forward, so a note appended only to it
  showed in that run's output but never reached disk. Fixed by resolving
  the cache's own object by identifier before appending.
• Running the check again on an already-flagged dataset duplicated the
  same sentence. Fixed by making the append idempotent (skip if the exact
  text is already present).

Both are covered by dedicated regression tests, plus manual stress
testing across multiple objects in one call, --refresh combined with
the flag, a first-ever call already using the flag, and a table with no
eligible shape.

balance correctly flagged as cumulative, purely from its shape in the data.

[marcociav-exmergo]

Looks good, thanks @emnarahmouni-exmergo