More consistent error types

[devmotion]

In a numerical estimation problem, I encountered numerical issues resulting in NaN parameters which then threw a DomainError when constructing a Normal distributions. So fine, so good - I could easily handle it by catching and handling the DomainError gracefully. Only later I noticed that similar numerical problems with Uniform will lead to an ArgumentError.

I think it's more convenient and more predictable for downstream users and packages if incorrect/infeasible parameter values always lead to DomainError.

Changes

• @check_args now requires every check to name the offending value(s) and always throws a DomainError carrying that value. The previous bare-condition and (cond, message) (ArgumentError) forms are removed; relational checks over several parameters (e.g. a < b) name them with a named tuple such as (; a, b) so downstream catchers can read the individual parameters off err.val.
• Remaining ad-hoc ArgumentErrors for infeasible parameters were converted to DomainError (or DimensionMismatch where a check is really about sizes/lengths, e.g. MatrixNormal, DiscreteNonParametric).

Merge with master / sparsity tracing

After merging master (which added global sparsity tracing through @check_args via _arg_checkable, #2064), the new named-tuple arguments needed to participate in the tracer-skip decision: a NamedTuple hit the default _arg_checkable(_) = true, so the relational checks still ran on primal-free tracers and errored.

This is fixed by recursing _arg_checkable through Tuple/NamedTuple, so such an argument is checkable only when all of its elements are. Global gradient tracing now works through the real-parameter constructors Uniform, Arcsine, LogUniform, Censored and TriangularDist. (DiscreteUniform and OrderStatistic use the same named-tuple form, but their checked parameters are integer-valued and can never be tracers, so global tracing does not apply to them; the recursion handles their named tuples consistently regardless.)

Because every multi-variable relational check now names all of its variables via a named tuple, the caveat noted in #2067 — that a cross-argument named condition could still slip through — no longer applies anywhere.

Added regression tests covering:

• global tracing through the affected constructors, including non-dense patterns (minimum/maximum/mode) to confirm dependencies are tracked rather than collapsing to a fully dense Jacobian;
• the check_args=false bypass under the local detector (primal present, so checks are not auto-skipped) with infeasible values, for both the scalar check (Normal's σ ≥ 0) and the named-tuple check (Uniform's a < b).

Closes #2067.

[codecov-commenter]

Codecov Report

❌ Patch coverage is 96.66667% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 86.63%. Comparing base (4e108fd) to head (a1dd254).

Files with missing lines	Patch %	Lines
src/utils.jl	83.33%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #2058      +/-   ##
==========================================
- Coverage   86.65%   86.63%   -0.02%     
==========================================
  Files         149      149              
  Lines        8925     8921       -4     
==========================================
- Hits         7734     7729       -5     
- Misses       1191     1192       +1     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.