Avoid redundant LUT buffer initialization

[GopalakrishnanN]

Description

• Allocate the temporary MLAS LUT buffer with std::make_unique_for_overwrite.
• Keep the existing full-buffer memset, preserving zero-initialized LUT data and padding while avoiding the allocation's redundant value-initialization pass.

Motivation and Context

MlasLutGemm currently value-initializes the temporary byte buffer and immediately zeroes the entire buffer again. The buffer ranges from a few KB during decode to multiple MB for representative prefill shapes, so the first write consumes memory bandwidth without affecting behavior.

Local Windows MSVC Release measurements used four alternating baseline/changed process runs with five repetitions each. The unaffected Dequant+SGEMM fallback served as a control.

Shape (M, N, K)	Before	After	Raw change	Control-normalized change
(128, 384, 1024)	478.3 us	454.3 us	-5.0%	-4.7%
(128, 1024, 4096)	3.506 ms	3.367 ms	-4.0%	-4.4%

Testing

• Built onnxruntime_mlas_benchmark and onnxruntime_mlas_test in Release mode.
• onnxruntime_mlas_test.exe --gtest_filter=SQLutGemmBlkBitWidth2BlkLen*.* (102 tests passed).
• lintrunner -a onnxruntime/core/mlas/lib/qlutgemm.cpp.

[Copilot]

Pull request overview

This PR optimizes the MLAS MlasLutGemm temporary LUT buffer allocation by avoiding redundant value-initialization while preserving the existing full-buffer zeroing that the implementation relies on.

Changes:

• Switch LUT scratch buffer allocation to std::make_unique_for_overwrite<int8_t[]> to skip default initialization.
• Keep the existing memset(..., 0, lut_buffer_size) to preserve the required zero-initialized LUT data/padding.