[TRITON]: Add fused GEMMs to optimize FF block

[willzhou-amd]

Adding fused GEMM + activation + gating options to reduce memory movement overhead in a FF block.

Changes:

• Add activation string parameter support for gemm_a16w16 to fuse an activation function with the GEMM.
• Add gemm_a16w16_gating.py which uses the first half of the (M, N) output along the N dimension as the gating activations and the second half as the layer activations. Produces a tensor of shape (M, N//2). This also includes an activation parameter for the gating activation.
• Write tests.
• Write a fully-fused E2E FF kernel. Will go into next PR.

Benchmarks on 350x:

Post-tuning:

Benchmarks on 300x:

Multikernel-ff is where we're at right now (e.g launching two Aiter Triton GEMMs).

[rahulbatra85]

Please add logging for the new op