perf(vindex): read search indexes through positional ranges

[JunRuiLee]

Purpose

This is PR 1/3 of a stacked vector-search performance series.

The regular/global-index vindex path currently buffers the complete index file
before search. For large indexes, a single query usually needs only metadata and
a subset of posting lists, so the full read adds unnecessary I/O and peak memory.

This change adds a safe positional-read foundation and switches the single-query
regular/global-index vindex path to it.

Scope boundary

• Multi-query global-index searches intentionally remain on the buffered path in
  this PR; perf(vindex): execute compatible queries through native batch search #646 replaces that fallback with native batch range reads.
• PK-vector ANN segment loading also remains buffered in this PR and is handled by
  perf(vector-search): execute PK-vector leaves concurrently #647.
• Lumina keeps its existing buffered read path.

Brief change log

• Add a SeekRead adapter backed by Paimon's asynchronous FileRead.
• Poll asynchronous storage reads on the caller's Tokio runtime instead of a
  second runtime.
• Return UnexpectedError instead of panicking when range-backed vindex search is
  driven outside a Tokio runtime.
• Run synchronous vindex and Lumina search work on a dedicated executor, avoiding
  block_in_place and host blocking-pool deadlocks.
• Honor global-index.thread-num as the per-query file fan-out while separating
  that logical limit from the process executor's physical OS-thread cap.
• Bound physical workers to max(32, 4 * available_parallelism) and reclaim
  growth workers after one idle minute, while retaining the CPU-sized baseline.
• Preserve strict sequential execution for a query when
  global-index.thread-num = 1.
• Validate requested ranges and report short reads explicitly.
• Coalesce nearby reads with bounded read amplification and share the range-I/O
  semaphore across cloned readers.
• Forward try_clone_reader and read capabilities through the type-erased vindex
  input adapter; the range reader deliberately reports storage-agnostic defaults
  because it accepts arbitrary range counts and the optimal window is backend-specific.
• Keep queued-task cancellation cheap and avoid blocking a current-thread Tokio
  runtime when an already-started search is cancelled.

Tests

• cargo test -p paimon vindex --lib — 118 passed
• cargo test -p paimon test_global_index_thread_num --lib — 2 passed
• cargo test -p paimon --test pk_vector_baseline_test — 8 passed
• cargo test -p paimon --test pk_vector_batch_test — 10 passed
• cargo test -p paimon --test pk_vector_java_fixture_test — 1 passed
• python3 scripts/dependencies.py verify
• python3 scripts/release_licenses.py --check — 21 generated legal files validated
• cargo fmt --all -- --check
• git diff --check

Coverage includes caller-runtime polling, a non-Tokio error path, current-thread runtimes, clone support,
short reads, bounded coalescing, physical worker clamping, idle worker reclamation,
and local filesystem reads when the host blocking pool has only one thread.

API and Format

No public API or storage-format changes.

Documentation

The existing global-index.thread-num API documentation now includes regular
vector global-index search. No user-facing storage or query API changes are required.

[JingsongLi]

+1