[jlblancoc]

Summary

Faster encoding and decoding of single-field clouds and lossless fields. The streams are byte-identical and the API is unchanged:

1. Bulk decodePoints() for FieldDecoderCopy. Lossless fields, and float fields without a resolution, were decoded through the generic per-point loop, with a virtual decode() call and a libc memcpy() call per value; perf showed about 90% of the time there. A single-field cloud now takes one memcpy. Interleaved points take a strided copy with a compile-time size.
2. Bulk encodePoints() for single-field clouds. I added a new virtual FieldEncoder::encodePoints(), whose default calls encode() once per point. The V4 codec uses it when a cloud has a single field. FieldEncoderCopy overrides it with a copy, and FieldEncoderFloat_Lossy with a loop that has no virtual call per point.
3. -O3 in non-Debug builds (GCC and Clang). RelWithDebInfo, the default of ROS/colcon builds, and distribution packages compile with -O2. There, the lossy float decoder runs about 1.5x slower.

Motivation: in MRPT, each point field is a separate array, so each one is encoded as its own single-field Cloudini stream (in 64k-point chunks, in parallel). That is about 2.8-3.5x faster than raw serialization for large maps. Maps with many lossless fields (normals, timestamps...) spend most of their time in these two paths.

Measurements

Times per value (or per point, for x/y/z) on a 262k-point LiDAR-like cloud, use_threads=false. Both builds are RelWithDebInfo (the default in ROS/colcon), GCC 13, AMD Ryzen 7 2700X. "1.4.1" is current main.

Decoding

	1.4.1	this PR
one lossless float field, ZSTD	7.5 ns	3.3 ns (2.3x), now bound by ZSTD
one lossless float field, LZ4	4.3 ns	0.25 ns (17x)
one lossless float field, no compression	4.5 ns	0.14 ns (32x)
x, y, z as 3 single-field lossy streams, ZSTD	21.3 ns	14.0 ns (1.5x)
x, y, z as 3 single-field lossy streams, LZ4	16.3 ns	10.3 ns (1.6x)
x, y, z interleaved (V6), ZSTD	15.2 ns	15.3 ns (unchanged)

Encoding

	1.4.1	this PR
one lossless float field, ZSTD	8.5 ns	4.8 ns (1.8x)
one lossless float field, no compression	4.5 ns	0.43 ns (10x)
one lossy float field, ZSTD	10.2 ns	7.5 ns (1.4x)

Testing

• New test FieldEncoders.SingleFieldBulkMatchesPerPoint. For contiguous and padded points, lossless and lossy, it checks that the bulk encoder writes the same bytes as encode() once per point. It also checks a whole-cloud round trip.
• I compared the streams of single-field clouds across sizes (1 to 300k points), point steps, lossless/lossy and NONE/ZSTD against 1.4.1: they are byte-identical (9.6 MB in total).
• All 58 tests pass.

[facontidavide]