[MLAS] Integrate prototype decomposable QGEMM

[JonathanC-ARM]

Summary

This change integrates Arm® KleidiAI™ prototype decomposable QGEMM kernels into the MLAS quantized GEMM path, enabling backend-selected acceleration for quantized CPU operators on supported Arm SME2
platforms.

The patch adds a new MLAS QGEMM override hook, threads MLAS_BACKEND_KERNEL_SELECTOR_CONFIG through quantized GEMM packing and execution call sites, and implements KleidiAI-backed {u8,i8} x {u8,i8} QGEMM
paths for both int32 accumulation output and fused float ScaleBias output.

Key Changes

• Updates the KleidiAI dependency to a prototype Arm GitLab commit containing decomposable GEMM kernel support.
• Adds an MLAS QGEMM backend override path alongside the existing SGEMM and DynamicQGEMM override mechanisms.
• Extends MlasGemmBatch, MlasGemm, and MlasGemmPackB to accept backend kernel selector configuration.
• Implements KleidiAI SME2 QGEMM support for:
  • u8 x u8 -> i32
  • u8 x i8 -> i32
  • i8 x u8 -> i32
  • i8 x i8 -> i32
  • Corresponding fused ScaleBias -> f32 variants
• Adds packed-B support for the KleidiAI QGEMM path by appending a backend-specific packed representation after the base MLAS packed buffer.
• Adds output processor introspection so backend overrides can detect and fuse MLAS_QGEMM_SCALE_BIAS_OUTPUT_PROCESSOR without RTTI.
• Propagates backend selector configuration through quantized CPU operators including:
  • MatMulInteger
  • QLinearMatMul
  • QLinearConv
  • ConvInteger
  • quantized attention
  • dynamic quantized matmul/LSTM
  • RNN helper GEMM paths
• Restores packed weight size metadata when shared prepacked buffers are reused by QAttention and QLinearConv.
• Adds a 1D unsigned ConvInteger im2col fast path.
• Optimizes single-GEMM DynamicQuantizeMatMul execution by calling MlasDynamicQGemm directly instead of the batch API.