BUG: fix convert_dtypes dropping values from sliced mixed-dtype DataFrames

[moktamd]

• closes BUG: Wrong dataframe data after convert_dtypes #64702
• Tests added and passed if fixing a bug or adding a new feature
• All code checks passed

Summary

After slicing a mixed-dtype DataFrame and calling convert_dtypes(), columns converted to ExtensionArray-backed types (e.g. Arrow strings) could silently lose values.

The root cause is in Block.convert_dtypes: after self.convert() splits the original object block into type-specific blocks, the loop still references self.shape[0] and self.dtype instead of blk.shape[0] and blk.dtype. When the original block had multiple rows (one per column in the consolidated object block), the converted single-row ExtensionBlock was unnecessarily passed through _split(), which sliced the 1-D backing array as if selecting a row from a 2-D array, truncating the data.

Fix: use blk.shape[0] and blk.dtype to reference the post-conversion block rather than the pre-conversion self.

Repro from the issue:

df = pd.DataFrame(data=[[1, "a"], [2, "b"], ["c", 3]], columns=["col1", "col2"])
df = df.loc[[0, 1]].copy()
df = df.convert_dtypes()
print(df["col2"].shape)  # (1,) — should be (2,)