feat binlog apply optimization

[shaohk]

Description

1. Support for ignoring binlog events that exceed chunk boundary values.
   If the value corresponding to the unique key of the binlog exceeds the maximum value of chunk iteration and is less than the maximum boundary value of the copy, it is ignored.

2. Support for binlog merge processing.
   When the columns of the unique key selected by chunk are int or float, the binlog is processed by map merging, and all delete operations are merged into one delete sql.All insert and update operations are merged into one replace sql. Then execute these sql in db transaction.

In case this PR introduced Go code changes:

• contributed code is using same conventions as original code
• script/cibuild returns with no formatting errors, build errors or unit test errors.

[dnovitski]

Correctness Issues Found + Fix PR

I reviewed this PR in detail and found several correctness bugs in the merge DML implementation. I've submitted a fix as PR #1687 (built on top of this branch).

Critical Issues

1. generateReplaceQuery reads live source state — REPLACE INTO ghost SELECT ... FROM source WHERE pk IN (...) reads the current source row, not the binlog event's row image. This is incorrect when:

   • Events are buffered during throttle (source may have changed since event time)
   • gh-ost is killed and restarted (replay uses current state, not original)
   • Cutover happens before all events drain

2. isIgnoreOverMaxChunkRangeEvent returns true when nil — Before the first row-copy chunk, MigrationIterationRangeMaxValues is nil, causing the comparison function to return true (ignore event). Early events are silently dropped.

3. Missing column conversion tokens — If using plain ? placeholders, timezone (convert_tz), enum (ELT), and JSON (convert(? using utf8mb4)) columns would be inserted without their required conversions.

Fix (PR #1687)

• Uses REPLACE INTO ghost (cols) VALUES (?, ?) with binlog after-image values
• Handles column conversions via BuildColumnsPreparedValues
• Fixes nil check to return false (don't ignore)
• Introduces mergedEntry struct for safe map value handling

Performance Results

The fix maintains the performance benefits:

• 10% faster general workload, 39% faster hot-spot dedup, 2.25× sustained throughput vs non-merge mode

See #1687 for full details and benchmarks.