[avison9]

Closes #3744

Rationale for this change

#3744 reports that after a spec-allowed promotion (int → long, float → double), every metadata
inspection API raises struct.error on tables that still have files written before the promotion. Bounds
are stored without a type tag, and older files keep the 4-byte encoding while the current schema type is
8 bytes.

This is already fixed on main by #3293 (released in 0.12.0): from_bytes(LongType/DoubleType) now falls
back to 4-byte decoding, and every bounds decode in inspect.py goes through it. The issue was reported
against 0.11.1, which predates that fix.

#3293 only added coverage for the metrics evaluators, though; the inspect path was fixed as a side effect
with no test pinning it. This PR adds that coverage so the crash can't silently come back, e.g. if
inspect moves to a different decode path (#4297).

tbl.append(pa.table({"qty": pa.array([1, 2], pa.int32())}))
with tbl.update_schema() as u:
    u.update_column("qty", field_type=LongType())

tbl.inspect.files()
# 0.11.1:  struct.error: unpack requires a buffer of 8 bytes
# main:    readable_metrics.qty -> lower_bound=1, upper_bound=2

Are these changes tested?

Test-only change. test_inspect_bounds_after_type_promotion writes with int/float, promotes to
long/double, and asserts correct decoded bounds from entries, files, data_files, all_files,
all_data_files, and the manifests() partition summaries.

Verified that it fails with the original struct.error when #3293's change to conversions.py is reverted,
and passes on main. Full unit suite: 4524 → 4525 passed, no new failures; make lint clean.

Are there any user-facing changes?

No.