Proposal: Add build-selectable SVE2-256 architecture support

The repository currently contains an SVE2-128 implementation under
include/sonic/internal/arch.

SVE2-256 is commonly used on Kunpeng950. Adding a separate SVE2-256
implementation would allow the code to take advantage of the wider vector
width and higher data-parallel throughput while continuing to use SVE2
instructions.

The implementation would be selected through a build option. The change is
approximately 1,200 lines and includes optimizations for
both serialization and deserialization.

In preliminary testing with Clang 16 and the existing benchmark testdata,
compared with the NEON build on the same Kunpeng950 system:

- Deserialization shows about 13%+ average CPU-time reduction, up to 35%.
- Serialization shows about 4% average CPU-time reduction, up to 15%.

Could you please confirm:

1. Is adding a separate SVE2-256 architecture directory acceptable?
2. Is the existing testdata sufficient for benchmark evaluation?
3. Is the NEON build on the same Kunpeng950 system an acceptable baseline?

Start by reading the existing SVE2-128 implementation under include/sonic/internal/arch and review how architecture selection is exposed through the build. Then inspect the existing benchmark testdata and benchmark results. Done would be an accepted build-selectable SVE2-256 implementation covering serialization and deserialization, with benchmark evaluation against the stated NEON baseline.

Written by the indexing model from the issue text.