Optimize Parser.raw to batch raws.before/between extraction in a single token scan

[alanzabihi]

Hypothesis

During AST construction, the parser processes tokens in several methods (decl, rule, atrule, other) that each independently handle whitespace/comment tokens. For a typical declaration, the token flow involves:

1. Collect tokens in other() until reaching a boundary
2. Call spacesAndCommentsFromEnd() — scans tokens backwards
3. Call spacesAndCommentsFromStart() — scans tokens forwards
4. Call raw() to extract the value — iterates tokens again

These three passes over the same token array are redundant. A single forward scan could extract before-spaces, the prop, between-spaces, the value, and after-spaces in one pass, populating all raws fields simultaneously.

Implement a single-pass token processor for the common declaration case:

// Single pass: scan tokens[0..n], partition into:
// before_spaces | prop_tokens | between_spaces | value_tokens | after_spaces
// Set all raws in one go

Editable surface

• lib/parser.js — single-pass token processing for common declaration/rule patterns

What's different from prior work

• Reduce Parser token array intermediate allocations in spacesAndCommentsFromEnd #48 replaced shift() with index tracking in spacesAndComments (neutral — V8 optimizes small array shift). This thesis eliminates the multiple scan passes entirely.
• Specialize Parser.decl() and Parser.rule() fast paths for common CSS patterns #21 specialized decl/rule fast paths but kept multi-pass scanning. This merges the passes.
• Flatten Parser.other() token scanning loop to reduce branch depth #31 flattened other() loop. Different code path.

Expected impact

METRIC_A improvement of 2-4ms (large file, thousands of declarations). METRIC_B improvement of 2-4ms. Combined 3-5ms.

[alanzabihi]

Polyresearch approval: thesis #73.