KleidiAI SVE2p1 halfgemm kernel integrated (#32670)
### Description
Integrates the Arm KleidiAI SVE2.1 FP16 HalfGemm kernel into MLAS.
Key changes:
- Updates KleidiAI from v1.20.0 to v1.30.0.
- Adds SVE2.1 runtime detection on Linux and Windows.
- Selects SME/SME2 first when available, otherwise SVE2.1.
- Adds SVE2.1 runtime and backend-native RHS packing.
- Records the packing thread’s SVE vector length in the native-packed
buffer.
- Prevalidates every native-packed batch entry against the caller’s
current vector length before
writing any output.
- Rejects mixed-VL batches without partially modifying their outputs.
- Validates worker-thread vector lengths before executing SVE2.1 tiles.
Work is replayed safely on
the validated caller thread when worker VLs differ.
- Rejects unsupported native-packed combinations such as a separate bias
or nonzero `ldb`.
- Adds overflow protection and rejects `K` values exceeding the kernel’s
supported range.
- Adds Arm copyright and SPDX licence declarations to the modified
implementation files.
### Motivation and Context
SVE vector length is configured per thread, and the SVE2.1 packed RHS
layout depends on that vector
length. A buffer packed by one thread cannot safely be consumed by a
thread using a different VL.
This change enables KleidiAI HalfGemm acceleration on SVE2.1 systems
without SME while preserving
SME/SME2 priority and preventing mismatched packed layouts from
producing incorrect or partially
modified output.
### Testing
Added coverage for:
- Exact and tail tile shapes.
- Runtime-packed and backend-native-packed RHS.
- Padded leading dimensions and guard regions.
- Single-threaded and threaded execution.
- Same-VL multi-entry batches.
- Batch entries packed with different vector lengths.
- Worker threads using a different VL from the packing/caller thread.
- Unsupported bias, transpose, leading-dimension, and `K`
configurations.
- Packed-buffer size and overflow handling.
A local Release KleidiAI build and CTest suite completed successfully.
SVE2.1-specific tests are
hardware-gated and skip on systems without the required SVE2.1
capabilities.
---------
Signed-off-by: melkap01 <melike.kaptan@arm.com>