Compose serialization into byte[] to avoid per-call Unsafe overhead - #12

…20)

* Add large-message serialize benchmark and non-array target identity sweep

LargeMessageBenchmark serializes the Pulsar topic-list shape from 600 B to
4.6 MB, a varint-dense repeated-int64 message at 2 KB and 8 KB, and a 2 MB
bytes payload into pooled direct buffers. NonArrayTargetIdentityTest checks
that writeTo() to direct, offset and multi-component composite targets is
byte-identical to the heap-array path for sizes swept byte by byte across
every plausible internal boundary, for built and parsed messages and across
repeated writes.

* Write direct buffers in place through their NIO view above 512 bytes

Since #12, writeTo() to a non-array buffer stages the whole message in a
heap byte[] scratch and bulk-copies it. Messages above SCRATCH_RETAIN_MAX
(1 MiB) never retain that scratch, so every write allocated a fresh
full-size array — multi-MB G1-humongous allocations that OOMed Pulsar's
proxy back-pressure test (apache/pulsar#26256, together with the clear()
retention fixed in #19). Below the cap the copy itself was still paid.

A single-region direct buffer exposes its memory as a java.nio.ByteBuffer
through ByteBuf.internalNioBuffer(). Absolute puts on a DirectByteBuffer
compile to a bounds check plus a jdk.internal.misc.Unsafe store — which,
unlike sun.misc.Unsafe, carries no JDK 24+ deprecation check — so the
message can be written in place: no scratch array and no bulk copy, at
any size. writeTo() now dispatches heap buffers in place through the
backing array (unchanged), single-region direct buffers larger than
NIO_WRITE_MIN (512 bytes) through the NIO view, and everything else
(small messages; composites and other buffers without a single NIO
region) through the scratch path as before. Above the threshold no
direct-buffer write touches the scratch, so it only grows past 512 bytes
for composite targets.

The threshold exists because the view's per-put cost is a fixed tax per
message while the copy it saves grows with size. Interleaved JMH on pooled
direct buffers (JDK 21/26): the view is 15-19% slower on the ~70-byte
varint-dense MessageMetadata, at parity on BaseCommand, 20% faster at
600 bytes, and 35-40% faster from 6 KB to 100 KB; on the 2 MB / 4.6 MB
cases it removes the per-write allocation (-70% / -55%) and matches the
per-field ByteBuf-API write-through of #18, which it replaces.

The field emitters are parameterized over the write sink (WriteSink.ARRAY
/ WriteSink.NIO): one emitter produces both _writeTo(byte[], int) and
_writeTo(ByteBuffer, int), differing only in the sink variable and in how
bulk data is copied out of a ByteBuf; every raw writer in LightProtoCodec
is overloaded for both sinks.

NonArrayTargetIdentityTest sweeps sizes byte by byte across every boundary
(64 B .. 1 MiB) on direct, offset and multi-component composite targets,
for built and parsed messages, repeated strings incl. non-ASCII, bytes
payloads, nested trees and the Pulsar BaseCommand shape. NioWriteTest
checks the routing flips exactly at NIO_WRITE_MIN, that composites keep
the scratch path, and (via ThreadMXBean.getThreadAllocatedBytes) that
5 writes of a 4.6 MB topic list or a 5 MB payload allocate less than a
quarter of one message. LargeMessageBenchmark covers 600 B .. 4.6 MB
topic lists, varint-dense 2 KB / 8 KB messages and a 2 MB payload.

Its last readers were the Unsafe raw writers that streamnative#12 replaced with the
byte[] writers. The NIO writers added later take the byte order from the
buffer itself, with nb.order().

* Remove the unused ByteBuf writers from LightProtoCodec

Generated code serializes through the raw byte[] and NIO writers, so
LightProtoCodec's ByteBuf writers (writeVarInt, writeVarInt64,
writeSignedVarInt*, writeFixedInt*, writeFloat, writeDouble and
writeString) were only called by tests. The codec is copied into every
generated package, along with these 78 unused lines.

LightProtoCodecTest and SegmentedByteBufTest now encode with the raw
writers that generated code uses. Their assertions are unchanged, so the
protobuf-java CodedInputStream checks now cover the production writers,
which had no direct unit tests.

* Remove the unused LITTLE_ENDIAN constant from LightProtoCodec

Its last readers were the Unsafe raw writers that #12 replaced with the
byte[] writers. The NIO writers added later take the byte order from the
buffer itself, with nb.order().