Fix row index miscalculation in ParquetLoader

[vini-fda]

Before submitting

• Was this discussed/agreed via a Github issue? (no need for typos and docs improvements)
• Did you read the contributor guideline, Pull Request section?
• Did you make sure to update the docs?
• Did you write any new necessary tests?

What does this PR do?

Fixes #809.

This replaces the uniform-size arithmetic with a per-chunk prefix-sum of row-group sizes, computed once when the ParquetFile is first opened, then use bisect.bisect_right(offsets, row_index) - 1 to locate the group and row_index - offsets[group] for the offset inside it.

The same num_rows_per_row_group value is also used in the cache-eviction check at

litData/src/litdata/streaming/item_loader.py

Line 749 in 1fdfad7

if read_count >= num_rows_per_row_group:

.

But that check needs the actual size of the current group (offsets[g+1] - offsets[g]), otherwise with uneven groups memory either leaks (never hits threshold) or is freed too early (forcing re-reads).

[codecov]

Codecov Report

❌ Patch coverage is 89.47368% with 2 lines in your changes missing coverage. Please review.
✅ Project coverage is 81%. Comparing base (1fdfad7) to head (3bdc6b5).

Additional details and impacted files

@@         Coverage Diff         @@
##           main   #810   +/-   ##
===================================
  Coverage    81%    81%           
===================================
  Files        54     54           
  Lines      7617   7630   +13     
===================================
+ Hits       6143   6157   +14     
+ Misses     1474   1473    -1     

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.