Refactor Superset model benchmarking tools to use Pydantic classes and save one json

[skhorasganiTT]

Ticket

#15435

Problem description

• The current benchmark infra for uploading data to Superset is using the deprecated method of generating 3 separate CSV files for run/measurement/environment data.
• The Llama3 t3k benchmark data was not being uploaded by the t3k demo tests

What's changed

• Refactored the model benchmarking tools to generate a single json file which is partially filled with run/measurement data during model tests, and afterwards completed with environment data by the CI test infra
• Refactored the benchmark data to be stored using the Pydantic classes implemented by the data science team so that data types are validated prior to json creation
• Updated the benchmark environment job in the Produce data for external analysis workflow to test completing a partial benchmark json
• Minor var cleanup in demo tests, modifying 'csv' to 'json'
• Enabled uploading of benchmark data for the llama3 T3K demo tests

Checklist

• Post commit CI passes
• Blackhole Post commit (if applicable)
• Model regression CI testing passes (if applicable)
• Device performance regression CI testing passes (if applicable)
• (For models and ops writers) Full new models tests passes
• New/Existing tests provide coverage for changes

[skhorasganiTT]

Produce data for external analysis test: https://github.com/tenstorrent/tt-metal/actions/runs/12818671201
Single-card demo tests: https://github.com/tenstorrent/tt-metal/actions/runs/12815226849
T3K demo tests: https://github.com/tenstorrent/tt-metal/actions/runs/12815926839
All post-commit tests: https://github.com/tenstorrent/tt-metal/actions/runs/12815232660

[mtairum]

Looks good from the llama3 demo side.

I'll be updating my local changes to use the new pydantic as well 👍

[tt-rkim]

I propose a decently big change in benchmarking_utils.py

Rest of the code is ok