Fix cross-language inconsistencies in gmd()

[Yicheng-Zhao1999]

Cross-language consistency fixes (Python)

Aligns global_macro_data/gmd.py with the R/Stata packages while reading the same
S3 dataset. Verified against live version 2026_03.

Coverage (this package = Python)

#	Finding	Needed in Python?	Changed in this PR?
1	All-missing rows not dropped as documented	Yes (drops only leading rows)	No — needs a product decision
2	Variable names case-sensitive	Yes	Yes — case-insensitive + canonical casing
3	list/current/load keywords case-sensitive	Yes	Yes — case-insensitive
4	Whitespace not trimmed	Yes (version)	Yes — version trimmed
5	raw with a non-boolean value	Yes (truthiness footgun)	No — pending decision (see below)
6	sources + a variable that source lacks	Yes (returned None)	Yes — now raises an error
7	Error precedence (version vs cite)	No (Python already validates version first)	No
—	print / network / fast	No (Python already has them)	No

What changed

• List of countries and list of indicators #2 variables="rgdp" matches and is normalized to canonical casing (rGDP);
  invalid names still error.
• Prepare 2.0.0 release #3 version="LIST", sources="LOAD", cite="LOAD", vars="LIST" now work.
• Align Python implementation with Stata semantics #4 version=" 2026_03 " is trimmed (country/variable tokens already were).
• Fix robustness bugs in gmd(): clean errors, flag coercion, year filtering, fetch hardening #6 sources + a missing variable now raises GMDCommandError instead of printing
  a message and returning None.

Not changed

• Update package to 1.0.0 #1 all-missing-row drop (Python/Stata drop only leading rows; R drops all).

Pending decision

• Wrap raw TypeError/RuntimeError in GMDCommandError #5 raw with a non-boolean value. Python uses plain truthiness, so any non-empty
  string enables raw — including raw="0" / raw="false" (a footgun). Target needs a
  decision: strict (only real booleans) vs lenient coercion ("true"/"1"/"yes" on,
  "false"/"0"/"no" off).