Reduce GCS QPS on ArrayRecordReader open via ReadLast. - #212

Previously, opening a `gs://` ArrayRecord shard issued 3 separate GCS HTTP GET
requests (5 when `index_storage_option` is `offloaded`):
1. An unbounded GET at offset 0 during `GcsReader` construction solely to obtain
   the object size (and then aborted mid-stream).
2. A GET via `reader->NewReader(size - 64KB)` to read the Riegeli postscript.
3. A GET via `reader->NewReader(footer_offset)` to read the ArrayRecord footer
   (and additional GETs in `OffloadedChunkOffset::operator[]` when
   `index_storage_option` is `offloaded`).

Open `riegeli::GcsReader` with `google::cloud::storage::ReadLast(1 MiB)` (using
a per-reader `google::cloud::storage::Client` instance to avoid cross-thread
`libcurl` handle pool contention and `Curl_peer` refcount data races) so the
initial HTTP GET retrieves the total object size (`Content-Range`) and
prefetches the last 1 MiB of the shard. In `ArrayRecordReaderBase::Initialize()`,
pull the prefetched tail into `reader`'s buffer (resetting the cursor to
`limit()` so `available() == 0` on close) and read the postscript and footer
(including offloaded chunk index lookups in `OffloadedChunkOffset::operator[]`)
directly from `reader`'s buffered tail via `ReadChunk` when
`footer_offset >= reader->start_pos()`, without spawning subordinate
`NewReader` HTTP streams.

Real-GCS integration benchmarks show 3.0×–5.0× fewer HTTP GETs on open (23.0× fewer on shards < 1 MiB), 2.0×–3.0× fewer TLS handshakes on open, 2.2×–4.7× faster shard initialization, and 1.13×–1.65× higher bulk read throughput with zero data races.

PiperOrigin-RevId: 992624347