Filter keywords from zlup identifier proptests and fix the unreachable orelse keyword alternative

[ciaranra]

Closes #419.

The reported bug

valid_identifiers_accepted generated raw [a-zA-Z_][a-zA-Z0-9_]* strings, so proptest eventually produces a keyword (minimal input fn), which the parser correctly refuses — the property asserted something false. parser_valid_function had the same latent defect in its function-name strategy (a generated name like if or true makes the source unparseable); the three remaining identifier-generating properties are guarded by if let Ok and merely skip keyword inputs, so they are left as-is.

Fix

The grammar is now the single source of truth for what counts as a keyword: zlup::parser::is_keyword matches a candidate against the grammar's own keyword rule (exact-length match), and both asserting properties prop_assume! keywords away through it. No keyword list is duplicated into the tests.

A real grammar bug this surfaced

Pinning is_keyword deterministically exposed that the keyword rule could never match orelse: pest's ordered choice matched the "or" alternative first, the trailing !(ASCII_ALPHANUMERIC | "_") lookahead then failed on e, and PEG parsing does not backtrack into a matched alternative. Because identifier is defined as !keyword ~ ..., the parser accepted orelse := 42; as an ordinary identifier assignment. The rule now lists "orelse" before "or". An audit of the full keyword list found no other prefix-shadowed pair (errdefer/error, union/unit, and try/true all diverge before either token ends). No example or test program used orelse as an identifier, and all five orelse-operator tests still pass.

keywords_are_rejected_as_identifiers pins the behavior in both directions: representative keywords (including orelse) are recognized and refused as identifiers, while near-keywords such as fnord, iffy, and returned remain valid.

Verification

Full cargo test -p zlup (666 unit + 175 proptest + CLI suites) green; cargo fmt --check, cold cargo clippy --locked -p zlup --all-targets -- -D warnings, and pre-commit run --all-files clean. The new pin test demonstrably fails against the unfixed grammar (that is how the orelse defect was found) and passes after the reorder.