fix: avoid memory-heavy Qlib preprocessing reshape

[ElhamDevelopmentStudio]

Summary

• Remove the global data_df.stack().unstack() operation from Qlib preprocessing.
• Process data directly by the named instrument index level.
• Preserve existing feature names, amount calculation, missing-value filtering, and minimum-history checks.
• Add regression coverage for sparse observations, missing values, and both supported MultiIndex orders.

Problem

Preprocessing large Qlib datasets could stall at:

data_df.stack().unstack(level=1)

This operation can materialize a large, dense datetime/field/instrument intermediate, causing significant memory amplification for sparse datasets.

The updated implementation groups the original dataframe by instrument and processes one instrument at a time without constructing the global intermediate.

Validation

• python -m pytest tests -q — 6 passed
• Focused preprocessing regression — 2 passed
• Python compilation check — passed
• git diff --check — passed
• Legacy and updated outputs compared across 500 instruments with sparse rows and missing values — exactly equal
• Synthetic sparse benchmark reduced the intermediate representation from 1,500,000 cells to the original 60,000-cell dataset shape

Issue

Fixes #161