feat: expose PostgreSQL SQLSTATE and diagnostic fields on errors

[schochastics]

fix #579

When a server-side error occurs, RPostgres currently surfaces only libpq's error message string and discards the SQLSTATE code and the other diagnostic fields, which forces users to regex-match localized message text to classify errors.

This PR captures those fields from the failing PGresult and raises a classed R condition that carries them: the condition gains the class vector c("RPostgres_error_<SQLSTATE>", "RPostgres_error", "error", "condition") and fields sqlstate, hint, detail, context, table, column, datatype, constraint, and schema (each NULL when libpq doesn't supply it).

Users can now dispatch on a specific code (tryCatch(..., RPostgres_error_FS002 = ...)) or read cnd$sqlstate directly, including for custom codes set via RAISE EXCEPTION ... USING ERRCODE. The original error message is preserved unchanged, and all errors—including connection-level ones with no result—gain the RPostgres_error parent class for uniform handling.

The condition is built in a small internal R helper invoked from C++; error sites that have a PGresult (step_run, prepare, and the COPY completion check) now route through it, while result-less paths keep their existing behavior aside from the new parent class. Includes tests covering undefined-table (42P01), syntax (42601), custom ERRCODE (FS002), and unique-violation (23505) cases.

[krlmlr]

Thanks.

• Could we also throw rlang errors with this approach?
• I'd like to include snapshot tests, needs local_edition(3) perhaps
• The approach differs from what duckdb is doing. I don't remember how I landed on the duckdb approach, perhaps just because I wanted rlang errors everywhere. I'm curious how and why this even works: R -> C++ -> R -> throwing and the resulting object survives. What does the captured call stack look like?

[krlmlr]

Needs pkgdown config: https://github.com/r-dbi/RPostgres/actions/runs/27250549490/job/80473863145?pr=581#step:19:606 .

[schochastics]

Ah sorry, missed the part about the rlang errors. I have another look

[krlmlr]

Red CI/CD.

[schochastics]

Claude says:
The remaining red checks are pre-existing CI issues, not from this PR — everything this PR touches is green

The two failing job types are unrelated to these changes:

1. All windows-latest jobs die in after-install:

   ankane/setup-postgres@v1/index.js:136
       throw `Postgres version not supported on Windows: ${postgresVersion}`;
   Postgres version not supported on Windows: 17
   

   after-install/action.yml defaults to pg-version 17, which ankane/setup-postgres@v1 can't install on Windows, so Postgres never starts and the job aborts before the build/tests.

2. ubuntu-24.04 (devel) errors on NOTEs (devel uses error-on: "note"), and there are two long-standing ones: Non-standard file/directory found at top level: 'configure.log' (left by the package's own configure script) and checking compilation flags in Makevars.

[krlmlr]

Yes, of course. Apologies, thanks for checking.