Add Llama3.1-8B benchmark with disabled collective matmul

[bvandermoon]

Description

Add a new Trillium benchmark that runs Llama3.1 with collective matmuls disabled. I ran this on v6e-8 and saw an improvement from ~350 TFLOP/s/device to ~410-420 TFLOP/s/device after this change.

This change is needed to support adding a reproducible recipe for v6e-8.

Tests

Ran this benchmark on v6e-8 using the following command:

python3 benchmarks/benchmark_runner.py xpk \
    --project=$PROJECT \
    --zone=$ZONE \
    --device_type=v6e-8 \
    --num_slices=1  \
    --cluster_name=${CLUSTER_NAME} \
    --base_output_directory=${OUTPUT_DIR} \
    --model_name="llama3_1_8b_8192_no_collective_matmul" \
    --libtpu_version=20241209 \
    --base_docker_image=maxtext_base_image

I got the perf described above. Also confirmed in the profile that the previous collective matmuls in the MLP layer are now gone:

Checklist

Before submitting this PR, please make sure (put X in square brackets):

• I have performed a self-review of my code.
• I have necessary comments in my code, particularly in hard-to-understand areas.
• I have run end-to-end tests tests and provided workload links above if applicable.
• I have made or will make corresponding changes to the doc if needed.

[gemini-code-assist]

Important

The terms of service for this installation has not been accepted. Please ask the Organization owners to visit the Gemini Code Assist Admin Console to sign it.