Minor updates to VLM Multimodal notebook

[xrdaukar]

Description

-- Refactor memory cleanup into a helper function and call it in more places.
-- Double the number of training steps (with the previous count I observed when adapter model still produces verbose outputs)
-- Retain some cells' outputs (they can serve as illustration)
-- More-consistent handling of variables .
-- Other minor changes

Related issues

Towards OPE-911

Before submitting

• This PR only changes documentation. (You can ignore the following checks in that case)
• Did you read the contributor guideline Pull Request guidelines?
• Did you link the issue(s) related to this PR in the section above?
• Did you add / update tests where needed?

Reviewers

At least one review from a member of oumi-ai/oumi-staff is required.