Wrap raw TypeError/RuntimeError in GMDCommandError

[karstenm1987]

Tracked upstream at Global-Macro-Database-Internal#414. Same filter as the recent Stata and R PRs: fix actual bugs, skip nice-to-haves.

Bugs

Both failure paths leaked internal exception types to users instead of the package's own GMDCommandError.

1. gmd(country=840) leaked TypeError: 'int' object is not iterable

Root cause: _tokens() did for item in value: with no type guard. Any scalar non-string (int, float, etc.) tripped Python's iteration protocol, not the package's own error surface. Same class of leak for variables=840, variables=1.5, etc.

2. get_available_versions() leaked RuntimeError when both mirrors were unreachable

gmd.py:333-344 caught the internal RuntimeError from _fetch_from, checked the on-disk cache and a local .dta fallback, and then bare-raise'd the internal error back to the user.

Fix

Two small changes:

1. _tokens() now rejects non-str / non-list / non-tuple inputs with GMDCommandError(code=198) and a message naming the actual type.
2. The terminal raise in get_available_versions() becomes a raise GMDCommandError(...) from exc with a message that tells the user what probably went wrong (no internet / transient outage).

Verified (live)

>>> gmd(country=840)
GMDCommandError: Expected a string or sequence of strings, got int.
>>> gmd(country=840.5)
GMDCommandError: Expected a string or sequence of strings, got float.
>>> gmd(variables=840)
GMDCommandError: Expected a string or sequence of strings, got int.

# Regressions:
>>> gmd(country="USA", variables="rGDP")          # DataFrame
>>> gmd(country=["USA","GBR"], variables="rGDP")  # DataFrame
>>> gmd(country=None, variables="rGDP")           # DataFrame
>>> gmd(country=["USA", 840], variables="rGDP")   # already caught; unchanged behavior

# Offline path (both mirrors failing, no local cache):
>>> get_available_versions()
GMDCommandError: Unable to fetch the list of available GMD versions.
                 Check your internet connection or raise an issue.

# Online sanity:
>>> get_available_versions()                       # list of versions

Out of scope

Other items in #414 are UX/consistency rather than correctness bugs and kept for separate PRs:

• raw/fast/iso truthy-string coercion (raw="no" still loads raw data) — same class as the Stata fast() UX wart, classified as "nice-to-have" in the Stata review.
• Whole-file download / no retry-backoff / no checksum / unversioned cache — design-level improvements.
• start_year / end_year — not implementing per maintainer preference.

Also note: the listed-but-404 versions on S3 (2025_01/03/05/06/08) are a release-pipeline issue and not fixable here.