perf: skip parseMap allocation and pool dynamicpb.Message

[yuzone]

Summary

Eliminate two major per-row heap allocations in the Flush hot path to reduce GC pressure and CPU overhead.

Changes

Direct processing of raw records

Process the map[interface{}]interface{} returned by Fluent Bit's msgpack decoder directly into proto binary, removing the parseMap deep copy into map[string]interface{}.

• Add rawMapToBinary / rawPopulateMessage / rawSetRepeatedField
• Add convertTimestampFieldsRaw to convert timestamps in-place on raw records
• Add []byte cases to all scalar converters to handle msgpack string values directly

Pool dynamicpb.Message

Reuse top-level dynamicpb.Message instances via sync.Pool keyed by proto FullName, avoiding a new allocation per row.

• Add getPooledMessage / putPooledMessage / marshalAndRelease
• Extract populateMessage from mapToMessage to allow filling a pre-allocated message
• putPooledMessage clears fields via Range+Clear, preserving internal map bucket memory

Remove dead code

Remove functions and tests that are no longer called after the above changes:

• jsonToBinary, parseMap, parseSlice
• convertTimestampFields, convertTimestampValue
• TestParseMap, TestParseSlice, TestConvertTimestampFields

Testing

• TestRawMapToBinary_* — raw map conversion including []byte values, nested structs, and repeated fields
• TestMessagePoolReuse* — verify no stale field leaks across pooled message reuse
• TestConvertTimestampFieldsRaw — timestamp conversion on raw records
• BenchmarkRawMapToBinary — benchmark for the raw path
• Existing Flush tests (TestFLBPluginFlushCtx, etc.) pass without modification