Allow train.max_steps for profiling runs

[MukeshK17]

Fixes #2187

What does this PR do?

Allows --train.max_steps to be passed via the CLI for pretraining runs.
Previously, this argument caused a validation error in pretrain.py.

The argument is now accepted and emits a warning clarifying that it is intended
for profiling, debugging, or minimal sanity-check runs.

Motivation

Users may want to run a very small number of training steps (e.g. max_steps=1)
to measure memory usage or execution time without committing to full pretraining.

Changes

• Removed train.max_steps from the unsupported argument list in pretrain.py
• Added a warning when train.max_steps is provided to clarify intended usage

Tests

• Core unit tests pass locally
• HF tokenizer tests requiring gated models fail locally without HF_TOKEN
• Optional dependency tests (datasets, bitsandbytes, lm_eval) not run locally
• CI is expected to cover gated and optional test paths

[bhimrazy]

LGTM 👍
@MukeshK17 , Could you also include a test for this? Maybe in test_cli or wherever it fits best. That would help ensure this behavior stays covered.

[MukeshK17]

Thanks! I’ll add a CLI-level test to ensure --train.max_steps is accepted and doesn’t raise a validation error. Will update the PR shortly.

[MukeshK17]

Added a CLI test in test_cli.py to cover --train.max_steps acceptance and warning behavior.
Thanks for the suggestion!