Adaptive Lossless Floating Point

[platypii]

Adds an ALP (Adaptive Lossless floating-Point) decoder for FLOAT and DOUBLE columns, per the parquet-format spec: https://github.com/apache/parquet-format/blob/master/AlpEncoding.md

• Page layout: 7 byte header, uint32 offset array, self-describing vectors (AlpInfo, ForInfo, packed values, exceptions).
• Decode is encoded * 10^factor * 10^-exponent with correctly rounded constants. FLOAT decodes in float32 arithmetic to match the reference implementation.
• Frame of reference is applied with wrapping int32/int64 arithmetic. DOUBLE uses a Number fast path when bit width <= 32 and FOR is a safe integer, BigInt otherwise.
• Exception values are copied as raw bytes so NaN payloads are preserved.

Tests:

• Unit tests build pages from the spec layout, including the worked example from the spec.
• alp_extended.zstd.parquet from parquet-testing is added to test files. All six ALP columns are compared bitwise against the PLAIN columns (9032 rows, vector sizes 32/1024/4096, NaN payloads, Inf, -0, subnormals, nulls, 64 bit FOR).