fix: reject non-finite floating-point default values

[huan233usc]

What

A float / double column default of NaN or ±Infinity passes SchemaField::Validate but cannot survive a metadata round-trip: the JSON serializer emits a non-finite number as null, which reads back as an absent default. The declared default is silently lost (and, because the key is present but null, the reader then errors on it).

How

ValidateDefault now rejects a non-finite floating default with a clear "value must be finite" error, so an unrepresentable default is caught up front rather than silently dropped on serialization.

The check uses std::isfinite, which is false for both NaN and ±Infinity — so a single condition covers all three cases (there is no separate IsNaN branch because isfinite already rejects NaN).

Testing

SchemaFieldTest.ValidateRejectsNonFiniteFloatingDefault explicitly covers NaN, +Inf and -Inf, and ValidateAcceptsFiniteFloatingDefault covers a normal value; verified fail-without / pass-with. Full schema_test passes (551 tests).

[evindj]

This looks good to me. For my own understanding how is the caller supposed to handle this error?

[zhjwpku]

You mentioned NaN in the PR description but doesn't seem to handle that, is that intentional?

[huan233usc]

It is handled by following std::isfinite returns false for both NaN and infinity, so a NaN default hits the same non_finite branch and is rejected. The test ValidateRejectsNonFiniteFloatingDefault covers NaN, +Inf and -Inf explicitly.

[zhjwpku]

I think java impl has the same issue, created a PR to strict the validation apache/iceberg#17106

[huan233usc]

@evindj thanks for the review! Validate() returns a Status, so the error propagates up through the schema / UpdateSchema construction path the same way other invalid-schema errors do — the caller gets a failed Result and never builds a schema carrying a default that could not be round-tripped through metadata.