Replace deprecated ParquetObjectReader with a hand-written AsyncFileReader

[jspaezp]

parquet 59 deprecates ParquetObjectReader. Upstream's replacement is not a
rename — you implement AsyncFileReader yourself
(apache/arrow-rs#10308).

Three #[allow(deprecated)] currently hold the line:

• rust/timscentroid/src/storage.rs — read_parquet_peaks
• rust/timscentroid/src/lazy/query.rs — ParquetQuerier::new_async, ParquetQuerier::query_async

What the replacement has to cover

AsyncFileReader requires get_bytes and get_metadata. The catch is
get_byte_ranges: its default implementation is serial —

for range in ranges {
    let data = self.get_bytes(range).await?;
    result.push(data);
}

ParquetObjectReader overrides it with ObjectStore::get_ranges, which fetches
concurrently. Taking the default would serialize every column-chunk fetch, so
the override is required, not optional. ParquetObjectReader also does metadata
suffix-prefetch and caching worth matching.

So the work is roughly:

• get_bytes → ObjectStore::get_range
• get_byte_ranges → ObjectStore::get_ranges (must stay concurrent)
• get_metadata → ParquetMetaDataReader with a suffix prefetch hint

Verifying it

Parquet reads here go through InstrumentedStore, whose counters now have test
coverage in rust/timscentroid/src/instrumentation.rs. A regression in fetch
batching shows up as a higher get_count for the same read, which is worth
asserting so a serial get_byte_ranges cannot land unnoticed.