fix: avoid internal errors for OneOf signature mismatches

[myandpr]

Which issue does this PR close?

• Closes Type signature matching errors are overly verbose and show internal details #20501.

Rationale for this change

After #18769, unsupported calls such as SUM(Boolean) started producing misleading internal errors.

The immediate problem is that TypeSignature::OneOf aggregates branch failures into an Internal error, which then becomes visible during planning together with internal TypeSignatureClass::... details.

This PR is intentionally narrower than #20070 and only fixes that TypeSignature::OneOf internal-error path.

What changes are included in this PR?

In datafusion/expr/src/type_coercion/functions.rs:

• keep the existing success path when any OneOf branch matches
• change the all-failed OneOf case to return Function '...' failed to match any signature as a planning error rather than an internal error

This does not restore the older Sum not supported for Boolean wording.

That older message came from a more function-specific path, and this PR intentionally does not try to infer a single branch-specific error from OneOf, since that can be misleading for overloaded functions with different arities or multiple valid type families.

In those cases, returning a branch-specific error can report the wrong arity or an overly specific type requirement that only reflects one overload.

Are these changes tested?

Yes.

Added unit tests covering:

• OneOf mismatches no longer returning Internal error
• OneOf arity mismatches also returning a planning error rather than an internal error

Updated sqllogictests for affected user-visible errors, including:

• sum(Boolean)
• nth_value(...)
• substr(...)
• generate_series(...)

Are there any user-facing changes?

Yes.

For failed TypeSignature::OneOf function calls, DataFusion now returns a normal planning error instead of an Internal error, while continuing to show the function call and candidate signatures.

[martin-g]

Suggested change

	Signature::coercible(
	vec![Coercion::new_exact(TypeSignatureClass::Decimal)],
	Volatility::Immutable,
	)
	.type_signature
	.clone(),
	Signature::coercible(
	vec![Coercion::new_exact(TypeSignatureClass::Duration)],
	Volatility::Immutable,
	)
	.type_signature
	.clone(),
	TypeSignature::Coercible(vec![Coercion::new_exact(TypeSignatureClass::Decimal)]),
	TypeSignature::Coercible(vec![Coercion::new_exact(TypeSignatureClass::Duration)]),

[myandpr]

updated to construct the TypeSignature directly in the test.

[myandpr]

fixed ci

[myandpr]

The current CI failures are due to an outdated clickbench.slt expectation rather than the OneOf changes in this PR.

The failing EXPLAIN for COUNT(DISTINCT ...) now gets optimized to ProjectionExec + PlaceholderRowExec because exact NDV can be derived from Parquet metadata.

That was fixed in #21050.

I have rebased this branch onto the latest main, so the updated expectation is now included here as well.