[GEMM] Fix functional issues with non-float16 dtype

[whitneywhtsang]

09-experimental-block-pointer.py tutorial is implemented for GEMM float16 dtype.
It can be executed as TRITON_INTEL_ENABLE_BLOCK_PTR=1 python 09-experimental-block-pointer.py.
This issue is to fix the functional issues with non-float16 dtype, e.g., int8, bfloat16, tf32.

---

Data type combinations of interest:

• bf16_bf16_fp32
• fp16_fp16_fp32
• i8_i8_i32 Minimal viable support for int8 on the block pointer path #1155
• tf32_tf32_f32 Minimal viable support for tf32 on the block pointer path #1172
• fp16_fp16_fp16 Minimal viable support for f16 dot product with f16 accumulator on the block pointer path #1211 Separate accumulator and result types #1258
• fp16_fp16_fp32_fp16 Separate accumulator and result types #1258
• bf16_bf16_fp32_bf16 👉 [GEMM] Fix functional issues with bfloat16 result format #1264
• bf16_bf16_bf16 Minimal viable support for bf16 dot product with bf16 accumulator on the block pointer path #1213 👉 [GEMM] Fix functional issues with bfloat16 accumulation #1265

[whitneywhtsang]

bfloat16 is handled in #1127, please continue investigation on the other dtypes.

[whitneywhtsang]

We can reference https://github.com/bashbaug/SimpleOpenCLSamples/tree/matrixperf/samples/99_matrixexperimentstf32 for tf32 support.