[TritonCPU] Prepare L1 matmul and elementwise kernels for benchmarking

[jopperm]

• Derived bench-cpu specs from bench-gpu (if they existed) or from the original KB config.
• Set block sizes in matmul kernels to AMX-friendly 32x32x32.
• Set assume_in_bounds kernel parameter to override tensor descriptor bounds checking where possible.
• Explore more block sizes for elementwise kernels.
• Persistent ReLU and GeLU kernels didn't perform well in initial testing; manually rewrote them as normal kernels.

[sandlbn]

LGTM !