parquet: Add an option to not parse the Page Index on each query

[progval]

Is your feature request related to a problem or challenge?

CREATE TABLE does not parse the Page Index, and SELECT does not cache it. This can make requests on large Parquet datasets take a significant time for queries that have a small number of results.

For example, with a simple SELECT int_column, other_int_column WHERE int_column=123456 on a table with 184 billion rows (so about 9 million Page Index items, given the default 20k page size)

output_rows=0, elapsed_compute=96ns, num_predicate_creation_errors=0, page_index_rows_filtered=0, predicate_evaluation_errors=0, row_groups_pruned_bloom_filter=21050, row_groups_matched_bloom_filter=0, file_open_errors=0, file_scan_errors=0, bytes_scanned=25023432248, row_groups_matched_statistics=21050, pushdown_rows_filtered=0, row_groups_pruned_statistics=173576, time_elapsed_scanning_total=16.763964ms, page_index_eval_time=3.153918ms, time_elapsed_scanning_until_data=16.745759ms, time_elapsed_processing=61.531313027s, time_elapsed_opening=96.012649352s, pushdown_eval_time=382ns

Describe the solution you'd like

Parse it once and for all, either on CREATE TABLE or lazily as SELECT queries read the files. (Note that in the case of partitioned tables, not all files may be read by the first SELECT)

Describe alternatives you've considered

https://github.com/apache/datafusion/blob/3b93cc952b889cec2364ad2490ae18ecddb3ca49/datafusion-examples/examples/advanced_parquet_index.rs

but it requires using the low-level API, and is not available through the SQL or Python interfaces.

Additional context

No response