Add eval_start_step flag and allow evaluation from step 0

[zxhe-sean]

Description

This PR adds full support for the eval_start_step configuration flag in MaxText to delay or control when interval-based evaluations trigger during pre-training, and updates the evaluation step check from step > start_step to step >= start_step.

Why is this change being made

1. Support for eval_start_step: The need of the delay is coming from the new validation requirement of MLPerf

The validation starts after N(GBS) = GBS*FLOOR(42+24576/GBS) samples - this gives us 2-3 spare evaluations before we hit the RCP min, and ~4 before we hit the RCP avg.

2. Support evaluation on step 0 / start_step (step >= start_step): The evaluation condition previously checked step > start_step, strictly preventing evaluations from firing on the very first step (step == start_step), such as when start_step = 0 and eval_interval = 1. Updating this to step >= start_step allows evaluation to fire on step 0 or the starting step when configured.

FIXES: b/535641549

Tests

• When setting eval_start_step=10, the log shows the eval started at step 10: https://fusion2.corp.google.com/invocations/f6ea1094-7d28-4c76-be75-f0bbaeb9ac5d/log

• The eval now can start from step 0: https://fusion2.corp.google.com/invocations/7cb76b6f-e36e-4ba7-ab17-aa90a9e1c6e6/log

Checklist

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

• I have performed a self-review of my code. For an optional AI review, add the gemini-review label.
• 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, including adding new documentation pages to the relevant Table of Contents (toctree directive) as explained in our documentation.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[RissyRan]

Thanks for the change!

[aireenmei]

LGTM!