build: Upgrade to unsloth 2026.3.3

[vivekkalyan]

Upgrade to unsloth==2026.3.3 and unsloth-zoo==2026.3.1

Problem

With newer unsloth versions, we were getting OOMs on certain training runs (which wasn't present in previous versions).

Solution

With older unsloth (2025.12.9), RL training mode effectively used model training defaults (model.for_training() with no explicit trainer gradient checkpoint arg path), which behaved memory-safe in ART.

With newer unsloth (2026.x), RL training mode reads trainer args and passes gradient_checkpointing explicitly into for_training(...), and this was reading TRL defaults (effectively False), which disabled gradient checkpointing and caused OOM.

So we are explicitly setting the default for gradient_checkpointing to be True