Improve out-of-box performance for GEMM kernels variants

[etiotto]

We have achieved good performance (relative to the XeTLA library) for a GEMM kernel (see http://benchmarks.glados.intel.com/d/1pXX4hUSz/microbenchmarks?orgId=1). Now is time to focus on improving performance of several variants of the GEMM workload:

1. GEMM + preOp (e.g. exp) applied to one input of tt.dot (https://github.com/intel/intel-xpu-backend-for-triton/blob/main/benchmarks/triton_kernels_benchmark/gemm_preop_exp_benchmark.py)
2. GEMM + GELU applied to tt.dot output (https://github.com/intel/intel-xpu-backend-for-triton/blob/main/benchmarks/triton_kernels_benchmark/gemm_postop_gelu_benchmark.py)
3. GEMM + matrix addition applied to tt.dot output (https://github.com/intel/intel-xpu-backend-for-triton/blob/main/benchmarks/triton_kernels_benchmark/gemm_postop_addmatrix_benchmark.py)

Work Items

• Bad performance for GEMM with cast operation on tt.dot operand #2346
• AllocateSharedMemoryPass has possibility to allocate SLM size greater than device max share memory #1716
• Failure to compile gemm_postop_addmatrix_benchmark.py with  #2378

[etiotto]

For GEMM + preOp (e.g. exp) applied to one input of tt.dot (https://github.com/intel/intel-xpu-backend-for-triton/blob/main/benchmarks/triton_kernels_benchmark/gemm_preop_exp_benchmark.py) PR #2346 improve performance by 16%.

[etiotto]

For GEMM + matrix add (postOp), PR #2400 improves performance from ~66TFlops to ~215TFlops for a 8Kx8Kx8K shape (other shapes also improve).

[etiotto]

This is done.