Refactor: Execute Maxtext inference as a module via python3 -m

Updated the recommended execution command for the inference script
from `python3 MaxText/decode.py` to `python3 -m MaxText.decode`.

This change utilizes Python's standard module execution mechanism (`-m`),
ensuring the script runs within the proper package context (`MaxText`).
This improves consistency and robustness of import resolution.

No changes were made to the source code itself.

PiperOrigin-RevId: 749207719

Author: Yijia Jin

benchmarks/mlperf/scripts/launch_microbenchmark.sh

@@ -12,7 +12,7 @@ export run_dir=${base_output_dir}/microbenchmark/${run_name}/${experiment_time}/
 echo "run_dir: ${run_dir}"
 gsutil cp ${config_file_path} ${run_dir}/
 
-python3 MaxText/inference_microbenchmark.py \
+python3 -m MaxText.inference_microbenchmark \
     ${config_file_path} \
     model_name=${model_name} \
     tokenizer_path=assets/tokenizer.llama2 \

benchmarks/mlperf/scripts/launch_server.sh

@@ -3,7 +3,7 @@ source ./${config}.sh
 export run_name=${model_name}_${tpu}_${attention}_ici_${ici_tensor_parallelism}-${ici_autoregressive_parallelism}_${reshape_q}_${quant_mode}_pbs${per_device_batch_size}_${compute_axis_order//,/}-${prefill_cache_axis_order//,/}-${ar_cache_axis_order//,/}
 
 cd /maxtext
-python3 MaxText/maxengine_server.py \
+python3 -m MaxText.maxengine_server \
     ${config_file_path} \
     model_name=${model_name} \
     tokenizer_path=assets/tokenizer.llama2 \

jetstream/tools/maxtext/model_ckpt_finetune_with_aqt.sh

@@ -50,7 +50,7 @@ export CONVERTED_CHECKPOINT=${MODEL_BUCKET}/${MODEL}/${MODEL_VARIATION}/${idx}/0
 # Fine tune the converted model checkpoints with AQT.
 export RUN_NAME=finetune_aqt_${idx}
 
-python3 MaxText/train.py \
+python3 -m MaxText.train \
 MaxText/configs/base.yml \
 run_name=${RUN_NAME} \
 base_output_directory=${BASE_OUTPUT_DIRECTORY} \