feat(no-undef): support eslint-style /* global */ comments

[bartlomieju]

The no-undef rule only recognized identifiers from its built-in globals
list, so legitimate ambient globals (declared elsewhere, injected by a host,
or specific to a non-Deno environment) were always reported as undefined with
no way to allow them. ESLint solves this with /* global foo */ directive
comments, but deno_lint had no equivalent, leaving the comment-based escape
hatch unavailable.

This teaches no-undef to read eslint-compatible /* global foo */ and
/* globals foo, bar */ block comments and treat the listed names as defined.
Writability hints (foo:writable, bar:readonly, and the legacy baz:true)
are parsed and ignored, since no-undef only cares whether a name is defined.
Matching is deliberately strict: only block comments count, and the keyword
must be exactly global or globals followed by whitespace, so identifiers
like globalThis and line comments are never mistaken for a directive.

Closes #1287

[bartlomieju]

Verdict: LGTM — correct and thoroughly tested. Recommend merge (rule-behavior change, so flagging rather than auto-merging; just needs green CI / a rebase).

Correctness — traced the parser against every test

• trim_start().strip_prefix("global") → optional strip_prefix('s') → require the next char is whitespace: this correctly accepts global/globals and rejects globalThing/globalThis (the whitespace guard is the key), matching the negative test.
• Comma-split with entry.split(':').next() correctly parses names and discards eslint writability hints (:writable/:readonly/legacy :true), including the spaced a: true form.
• Multiline directives work (/* global\n a,\n b\n*/) because the split operates on the raw text and each entry is trimmed.
• Line comments are ignored (CommentKind::Block guard), and undeclared names still error — both covered by tests.
• declared_globals.contains(&*ident.sym) looks up &str against HashSet<String> via Borrow, so no per-identifier allocation. The pre-pass collect_declared_globals borrows comments before the mutable context borrow — ordering is correct.

Test coverage is excellent: positive cases (all forms + hints + multiline), and negatives for undeclared names, line comments, and the globalThing false-directive.

Minor (non-blocking)

• Space-separated names like /* global a b */ become a single "a b" entry, which simply never matches any identifier (harmless). ESLint primarily splits on commas too, so this is close enough; if you want full parity you could also split entries on whitespace.

Perf: good — one pass over comments; O(1) set lookup per identifier. (The existing linear GLOBALS.iter().any() scan is unchanged and out of scope here.)

Housekeeping: mergeable: UNKNOWN → rebase; then it's ready. As a no-undef behavior change it's a rule-behavior tweak, but a well-scoped, opt-in-by-comment one.

No changes requested. Good to merge once CI is green.