cosmic-ray's mutation run over the diff found one surviving mutant:
`half = batch.num_rows // 2` mutated to `/`. A real pyarrow
RecordBatch.slice() silently truncates a float offset/length to the
same int, so no test built on real batches can observe the
difference. Add a duck-typed fake batch that records the exact
slice() call arguments and asserts they stay int, which does
distinguish the two operators.

Signed-off-by: dbbvitor <vitor.diniz@gympass.com>

Extract to_arrow()'s metrics/audit finally-block into a reusable
_emit_offline_store_request_metrics helper, and give every RetrievalJob a
default to_arrow_reader() that wraps to_arrow(). This keeps every existing
offline store working unchanged while giving stores that can page results
natively (Trino, next) a hook to override without duplicating metrics code.

Signed-off-by: dbbvitor <vitor.diniz@gympass.com>

Add Query.start()/iterate_pages() to page through cursor.fetchmany()
instead of fetchall(), and override TrinoRetrievalJob.to_arrow_reader()
to consume them lazily. Jobs carrying on-demand feature views still use
the materialized to_arrow() path, since ODFVs need the full table.

Also fix trino_to_pa_value_type() to map "... with time zone" to a
tz-aware pa.timestamp("us", tz="UTC") instead of a tz-naive one, since
the cursor returns tz-aware datetimes for those columns.

Signed-off-by: dbbvitor <vitor.diniz@gympass.com>

Switch _execute_read_api's three read APIs, and do_exchange's
get_historical_features branch, from to_arrow() to to_arrow_reader().
The default to_arrow_reader() makes this a no-op for every non-Trino
store; Trino now streams instead of materializing the full result
before the existing _bounded_reader (PR A) re-batches it for Flight.

SQL errors still surface before the stream starts, since Query.start()
executes eagerly. Cursor/temp-table cleanup on an early client
disconnect is best-effort: RecordBatchReader.close() does not close
the underlying generator.

Signed-off-by: dbbvitor <vitor.diniz@gympass.com>

Cosmic-ray found real gaps: the elapsed-time subtraction in both the
metrics helper and to_arrow_reader() had no test pinning its exact
value, so add/mul/mod/floordiv mutants survived; _stringify_complex's
depth==0 check and its depth-1 recursion had no negative-depth or
multi-level test to distinguish them from <=0 or >>1/%1/^1; and
_complex_column_depth's trailing slice needed three levels of array
nesting before an off-by-one became observable.

Also drop the dead preserve_index=False kwarg from
pyarrow.Table.from_pandas() in the Trino reader: since we always pass
an explicit schema with no index field, pyarrow ignores preserve_index
entirely for the table's data and public schema (verified empirically),
so the argument was an untestable no-op rather than a missing test case.

Signed-off-by: dbbvitor <vitor.diniz@gympass.com>

Update the RetrievalJob functionality matrix in overview.md to match
the per-store trino.md page: Trino now streams via to_arrow_reader().

Signed-off-by: dbbvitor <vitor.diniz@gympass.com>