Consolidate transient-keyword lexing: context-blind strip can corrupt string/comment text in the AST

[apbendi]

Background

Because solang-parser (even at latest upstream) cannot parse the transient keyword, scopelint handles it in two independent layers:

1. src/parser.rs — on parse failure, a regex blanks every \btransient\b with same-length whitespace and re-parses. This regex is context-blind: it also blanks occurrences inside string literals and comments.
2. src/check/validators/variable_names.rs — the validator re-scans the original source with a careful comment- and string-aware state machine to decide whether a state variable is transient.

Problem 1: silent corruption of string/comment text

When the fallback triggers (i.e. the file really uses transient storage), any transient token inside a string literal or comment is also blanked. The AST's string values and the parsed comment text then contain nine spaces where the word was. Downstream consumers that read that text — the EIP-712 typehash validator reads string literals, and spec/inline-config read comment text — would see corrupted content. This requires a file that both uses transient storage and mentions the word in a load-bearing string or comment, so it is unlikely in practice, but the failure would be silent when it hits.

Problem 2: three lexers for one job

• declaration_start_before_name and contains_keyword_before_name in variable_names.rs are ~150 nearly identical lines of comment/string-aware scanning.
• The parser.rs regex is a third, cruder implementation of "find transient tokens".

(The existing CommentState helper in comments.rs is not string-aware, so it cannot replace them as-is.)

Proposed fix

Do one careful, string/comment-aware lexing pass in sanitize() (in parser.rs) that:

1. blanks only real keyword occurrences (skipping strings and comments), fixing Problem 1; and
2. records the byte offsets of the blanked tokens into Parsed (e.g. transient_token_offsets: Vec<usize>).

The validator then determines "is this state variable transient" with a simple range query — is there a recorded offset between the declaration's start and the variable name — deleting the duplicated state machines in variable_names.rs (roughly half the code added in #64) and removing the per-variable O(file size) rescan as a bonus.

Existing unit tests in variable_names.rs (comment/same-line/ignore-directive cases) already pin the expected behavior, so the refactor can be validated against them.