[python] Support selected-key projection for shared-shredding MAP

[XiaoHongbo-Hope]

Purpose

Support literal-key projection for top-level MAP<STRING, ...> columns through the existing with_projection() API:

read_builder.with_projection([
    "id",
    "attributes['key.with.dots']",
])

Bracket syntax follows Spark MAP access, keeps dot notation for nested ROW fields, and avoids adding another ReadBuilder method. Exact top-level field names take precedence.

Behavior

• Shared-shredding files read only required physical columns and overflow.
• Other layouts and keys not safely encodable in the internal metadata fall back to reading the full MAP.
• Parquet, ORC, column rename, value type evolution, and data-evolution reads are supported.
• Full-MAP projection and MAP aggregation retain the existing behavior.
• Internally, PyPaimon reuses the Java/Spark temporary ROW and field-description protocol; it is not persisted.
• No new method or table option is introduced.

Tests

• 115 targeted Python tests pass.
• Java-written Parquet/ORC fixture tests pass.
• Python 3.6 syntax and git diff --check pass.