[ML] Fix BatchLoader::CreateTrainingBatches resetting wrong secondary buffer

[siliataider]

This Pull request:

Changes or fixes:

test14_big_data in rbatchgenerator_completeness was sporadically failing across multiple platforms with different errors depending on which random parameters were drawn (for entries_in_rdf, chunk_size and batch_size):

• wrong batch shape at the remainder batch
• missing entries

Both failures are caused by a small bug in RBatchLoader::CreateTrainingBatches, which, if we had a overflow in fPrimaryLeftoverTrainingBatch, was correctly promoting fSecondaryLeftoverTrainingBatch to primary but incorrectly resetting fSecondaryLeftoverValidationBatch afterwards (instead of fSecondaryLeftoverTrainingBatch), meaning that the training secondary leftover buffer was never cleared and with certain random combinations of parameters would cause issues.

We don't see this in master because the RBatchLoader was refactored in #20998

Checklist:

• tested changes locally

[github-actions]

Test Results

    19 files      19 suites   2d 20h 25m 54s ⏱️
 3 713 tests  3 713 ✅ 0 💤 0 ❌
68 832 runs  68 832 ✅ 0 💤 0 ❌

Results for commit 2a239f0.