fix(object_store): make get_ranges coalescing configurable

[Xuanwo]

Which issue does this PR close?

Follow-up to #7380 and #7383.

Rationale for this change

object_store_opendal::get_ranges currently executes one reader.read(range) for every requested range. The reads are buffered concurrently, but nearby ranges still cause separate backend reads.

Reader::fetch already provides range coalescing and concurrent execution. However, the earlier HDFS investigation in #7380 found that unconditional 1 MiB coalescing could approximately double cluster memory usage because small returned Bytes slices retained large coalesced backing buffers. The integration therefore needs a configurable coalescing threshold instead of choosing between always coalescing and never coalescing.

What changes are included in this PR?

• Route get_ranges through Reader::fetch with the existing concurrency of 8.
• Default the coalescing gap to object_store::OBJECT_STORE_COALESCE_DEFAULT (1 MiB).
• Add OpendalStore::with_get_ranges_gap so users can tune the threshold or set it to 0.
• Make gap = 0 a valid OpenDAL reader setting. It disables merging ranges separated by bytes while still merging overlapping or adjacent ranges.
• Add regression coverage for the default, disabled, and custom gap behaviors.

Are there any user-facing changes?

Yes. get_ranges now coalesces nearby ranges by default. HDFS users whose workloads are sensitive to retained coalesced buffers can construct the store with .with_get_ranges_gap(0) or choose an intermediate threshold.

Validation

• cargo test --manifest-path integrations/object_store/Cargo.toml
• cargo clippy --manifest-path integrations/object_store/Cargo.toml --all-targets --all-features -- -D warnings
• cargo test -p opendal-core -p opendal-layer-timeout test_merge_ranges
• cargo clippy -p opendal-core -p opendal-layer-timeout --all-targets -- -D warnings
• ./scripts/workspace.py cargo fmt -- --check

A real HDFS/DataFusion Comet cluster retest is pending.

AI Usage Statement

OpenAI Codex (GPT-5) assisted with implementation, tests, and PR preparation.

[Xuanwo]

@comphead, could you please retest this PR with the same HDFS/DataFusion Comet workload from #7380 and #7383?

Please compare at least:

1. The default 1 MiB gap.
2. .with_get_ranges_gap(0) to preserve exact-range reads.
3. An intermediate gap if practical for the workload.

The most useful results would be task wall-clock time and cluster memory usage. I am keeping the PR in draft until we understand whether the configurable gap avoids the earlier memory regression while improving the HDFS read path.