chore(dataset): flush the prior handle in _update_inplace (optional hardening)

[MAfarrag]

Note: tracking issue — NOT implemented in PR #798 (won't-fix as a bug; optional hardening only).

Context

Architecture review ARC-7 (P2→P3, downgraded — not reproduced as a bug). _update_inplace swaps self._raster
via __dict__.update(new.__dict__) without flushing/closing the prior handle.

Problem / Current Behaviour

The feared "file stays locked on Windows" outcome does not occur: a probe showed a weakref to the old handle
becomes None and os.remove(old_file) succeeds immediately after the swap — CPython refcounting drops the last
reference and SWIG's finalizer calls GDALClose. Engines hold only a weakref.proxy to the Dataset, not the
raster. Nothing is broken today.

Affected locations

File	Symbol	Notes
src/pyramids/dataset/dataset.py	_update_inplace	prior _raster not explicitly flushed/closed

Proposed Solution

Optional hardening only: capture prev = self._raster and prev.FlushCache() if prev is not src (matching the
explicit flush in ops/io.py _create_copy_and_reopen). Guards a future lingering reference / non-refcounting
interpreter.

Out of Scope

Any change premised on a current bug (there is none on CPython).

Effort Estimate

Size: S — but low value; kept as a tracked optional-hardening note.

Definition of Done

• _update_inplace flushes the prior handle when it differs from the incoming src
• No behavior change on CPython; suite green