trainingsample v0.3.0 for performance improvement on huge datasets

[bghira]

This pull request updates the required version range for the trainingsample dependency from >=0.2.18 to >=0.3.0,<0.4.0 and adjusts the related test to match this new requirement.

Dependency updates:

• Updated the trainingsample dependency in setup.py to require versions >=0.3.0 and <0.4.0 instead of the previous >=0.2.18.

Test updates:

• Modified the test in tests/test_setup_platform_dependencies.py to check for the new trainingsample version range and ensure the old version requirement is not present.

[bghira]

the improved tsr build brings total wall clock time on huge (2 million samples) datasets down from 4 hours to about 1 hour across 4x H200 and even better performance on gaming CPU class systems (eg. RTX 6000 Pro, RTX 5090, 4090) thanks to higher clock rates and boost freq