[Inductor][CPP] Add transposed B matrix support for CppMicroGemmFP32Vec (pytorch#147068)

* Add transposed B support for CppMicroGemmFP32Vec.
* Add support for cases where N is not divisible by `block_n`.
Expand CppMicroGemmFP32Vec to generate gemm kernel that supports transposed B and N of arbitrary size.
This is the basis for pytorch#147069 to get better performance.

Pull Request resolved: pytorch#147068
Approved by: https://github.com/leslie-fang-intel, https://github.com/jansel, https://github.com/jgong5

Author: CaoE

aten/src/ATen/cpu/vec/vec256/vec256_16bit_float.h

@@ -208,6 +208,12 @@ static_assert(
       return _mm256_loadu_si256(reinterpret_cast<const __m256i*>(ptr));
 
     __at_align__ int16_t tmp_values[size()];
+#ifndef __msvc_cl__
+#pragma unroll
+#endif
+    for (const auto i : c10::irange(count, size())) {
+      tmp_values[i] = 0;
+    }
     std::memcpy(tmp_values, ptr, count * sizeof(int16_t));
     return _mm256_loadu_si256(reinterpret_cast<const __m256i*>(tmp_values));
   }