Validate coerce int96 config 17498

[AlyAbdelmoneim]

Which issue does this PR close?

• Part of The config values aren't checked #17498.

Rationale for this change

Currently, invalid coerce_int96 values (e.g., "invalid") are accepted at SET time and only fail later when reading Parquet files. This PR adds early validation so invalid values are rejected immediately with clear error messages, following the same pattern as DFWriterVersion.

What changes are included in this PR?

• Add DFTimeUnit enum with FromStr, Display, ConfigField implementations
• Change ParquetOptions.coerce_int96 from String/Option<String> to Option<DFTimeUnit>
• Remove parse_coerce_int96_string as a separate runtime parser; validation now happens at config time
• Update proto conversions to validate during deserialization
• Update ParquetFormat::with_coerce_int96 and related usage to accept Option<DFTimeUnit>
• Add test for early validation

Are these changes tested?

Yes. Added test_parquet_coerce_int96_validation that verifies:

• Valid values ("ns", "us", "ms", "s") are accepted
• Invalid values ("invalid") are rejected immediately at SET time
• Error messages are clear and helpful

Are there any user-facing changes?

Not exactly. Invalid coerce_int96 values now error immediately when set via SET command or proto deserialization, instead of failing later during Parquet file reading. This provides better error messages and earlier feedback.

[AlyAbdelmoneim]

Hi @Jefffrey, I’ve been tied up with final exams, so I haven’t made progress on issue #17498 over the past few weeks. I’ll be resuming work on it in the coming days. Please review when you have a chance.

[Jefffrey]

If we treat None as defaulting to ns then perhaps we should remove the Option and just make it directly a DFTimeUnit?

[AlyAbdelmoneim]

yes I will do that

[AlyAbdelmoneim]

should I change CoerceInt96Opt too ?

[Jefffrey]

Hmm perhaps I was mislead by these test comments; I'm not certain that None does default to ns, especially considering theres some test failures now 🤔

Might need more investigation before proceeding with this change (of removing the Option)

[AlyAbdelmoneim]

The reason I initially proceeded with removing Option was this part of the docstring in config.rs:

/// (reading) If set, parquet reader will read columns of
/// physical type int96 as originating from a different resolution
/// than nanosecond.

I interpreted this as implying that the default behavior is effectively nanoseconds anyway, so making the field non-optional with a default of DFTimeUnit::Nanosecond seemed safe.

After reviewing the failing tests more closely, it’s clear that None does not behave the same as explicitly setting ns, and that None really means “don’t apply coercion at all”.

I’ll revert this change and keep Option<DFTimeUnit> to preserve the existing semantics.

Does that sound like the right approach to you?

[Jefffrey]

Yes that sounds good. And perhaps could improve the docstring for the config too, since it seems we were both mislead by what being unset means 😅

[AlyAbdelmoneim]

sure 😅

[comphead]

@mbutrovich FYI