optimize HBase 1.x request-path CPU usage

[hhlgt]

Summary

This PR optimizes the request and response hot paths of the HBase 1.x client.

It reduces CPU usage, byte-array copying, and temporary object allocation for
Get, Batch Get, Scan, and Put operations while preserving existing HBase client
semantics and compatibility paths.

Changes

Read path

• Consume compact K/Q/T/V result batches provided by obkv-table-client-java.
• Avoid materializing intermediate per-field ObObj values when compact results
  are available.
• Introduce OHBaseResultCell, a lightweight immutable Cell implementation
  backed directly by decoded field arrays.
• Optimize point Get result assembly by:
  • validating the returned row key once
  • reusing the canonical row-key array
  • avoiding temporary KeyValue objects and row-key clones
• Optimize closestRowBefore by retaining only cells belonging to the greatest
  returned row key without cloning row keys from temporary cells.
• Consume compact Scan rows directly and construct HBase 1.x-compatible
  KeyValue results without rebuilding generic rows.
• Avoid intermediate family/qualifier arrays for TableGroup
  family\0qualifier columns.
• Preallocate result containers and reuse a static HBase cell comparator.

Batch Get

• Add a dedicated result-processing path for pure Get batches.
• Preserve request/result ordering without maintaining an unnecessary
  per-operation result mapping.
• Consume compact K/Q/T/V batches directly.
• Preserve empty-result and per-operation exception behavior.
• Validate malformed result counts and K/Q/T/V payloads.
• Keep generic ObObj decoding as a compatibility fallback.

Put path

• When autoFlush is enabled, execute synchronous Puts directly without passing
  through the BufferedMutator queue, heap-size accounting, temporary linked
  lists, and an additional flush cycle.
• Flush existing buffered mutations before entering the direct path to preserve
  write ordering.
• Fall back to the buffered path if pending buffered state cannot be drained.
• Build Put V2 requests using compact cell arrays.
• Preallocate row, column-family, and cell containers.
• Reuse qualifier/value backing-array regions on the synchronous autoFlush path.
• Clone qualifier/value data on delayed-flush paths where the request outlives
  the current call.
• Avoid duplicate value cloning when encoding cells with TTL.
• Validate Put mutations using the original family/cell map without
  materializing deprecated compatibility structures.
• Preserve the existing server-version gate for multi-column-family writes.

Default behavior

The optimizations are applied automatically when the request shape and lifetime
make them safe. No new public tuning switches are introduced.

In particular:

• compact Batch Get decoding is selected automatically for eligible HBase
  requests
• compact Put encoding is enabled by default
• zero-copy Put cell handling is limited to synchronous autoFlush requests
• buffered writes continue to use owned copies
• Get uses lightweight result cells while Scan preserves the HBase 1.x
  KeyValue result contract

Compatibility

• Preserve the existing HBase 1.x public APIs.
• Preserve Get, Batch Get, Scan, and Put result semantics.
• Preserve TableGroup and single-column-family behavior.
• Preserve existence-only and empty-result handling.
• Preserve write ordering between buffered and direct Puts.
• Keep legacy decoding paths when compact results are unavailable.

Dependency

• Upgrade obkv-table-client from 2.4.0 to 2.4.1-SNAPSHOT.
• This PR depends on the corresponding optimization PR in
  obkv-table-client-java.

Dependent PR: oceanbase/obkv-table-client-java#442

Tests

Added 47 focused unit test cases covering:

• compact point Get and closestRowBefore processing
• compact Batch Get result conversion and ordering
• compact Scan result consumption
• TableGroup family/qualifier range handling
• synchronous autoFlush direct Put execution
• preservation of pending buffered-write ordering
• compact Put encoding with and without TTL
• zero-copy and sliced cell-array handling
• delayed-flush cell cloning
• Put validation and maximum cell-size boundaries
• lightweight result-cell compatibility
• malformed and empty result handling

[hhlgt]

AI Coding Review

审查范围：
Base: origin/master
Head: optimize/client-cpu-performance
Commits: 4
Target: HBase 1.x

[P1] BufferedMutatorParams.maxKeyValueSize 被静默忽略

位置：
src/main/java/com/alipay/oceanbase/hbase/util/OHBufferedMutatorImpl.java:173
branch: optimize/client-cpu-performance
OHBufferedMutatorImpl 仍然在构造函数中读取并保存：
this.maxKeyValueSize = params.getMaxKeyValueSize() ...
但 Put 校验被改成：
ohTable.validatePutMutation((Put) mt);
该方法使用的是 OHTable.maxKeyValueSize，而不是当前 BufferedMutatorParams 设置的 maxKeyValueSize。因此 mutator 字段虽然存在，却不再被使用。

影响：

• 用户为单个 BufferedMutator 设置更小限制时，超限 Cell 可能被接受并发送。
• 用户设置更大限制时，合法 Cell 可能仍被 OHTable 默认限制拒绝。
• 破坏 HBase BufferedMutatorParams.maxKeyValueSize 的既有参数语义。
• 建议增加带显式限制的校验入口：
  ohTable.validatePutMutation((Put) mt, maxKeyValueSize);
  或者保留：
  OHTable.validatePut((Put) mt, maxKeyValueSize);
  再单独执行现有的单 CF/多 CF 校验。

建议增加两个回归用例：
Table 默认值较大、Mutator 参数较小时，超出 Mutator 限制必须失败。
Table 默认值较小、Mutator 参数较大时，应按 Mutator 参数进行判断。

依赖风险

pom.xml 将依赖升级为：
<table.client.version>2.4.1-SNAPSHOT</table.client.version>
该 PR 依赖前一个 Table Client PR。合入前需要确保：
2.4.1-SNAPSHOT 已发布到 CI 可访问的 snapshot 仓库；或者
Table Client 发布正式版本后改为正式版本。

Review 结论

合入前需要修复 BufferedMutatorParams.maxKeyValueSize 参数失效问题，并确认 Table Client snapshot 的发布顺序。

[WeiXinChan]

LGTM